A radio direction finding anomaly detection method based on image segmentation

Abstract

This invention discloses a radio direction finding anomaly detection method based on image segmentation, comprising: acquiring real radio data; constructing a multi-signal classification spatial spectrum estimation function to extract the azimuth angle of the UAV signal; reassembling the azimuth angle data slices into a two-dimensional matrix set, calculating the true azimuth angle of the UAV, and constructing a training dataset; constructing a radio direction finding anomaly detection network based on an encoder and decoder structure; determining the optimization objective of the radio direction finding anomaly detection network and adjusting the loss function used for the network parameters; training the radio direction finding anomaly detection network using the training dataset; inputting test data into the anomaly detection network and outputting pixel-level azimuth angle prediction results. This invention achieves higher anomaly removal accuracy than existing methods, has strong transferability, can adapt to new radio azimuth angle acquisition environments, and has strong network interpretability, enabling better application in practical communication scenarios.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a radio direction finding anomaly detection method based on image segmentation. Background Technology

[0002] Anomaly detection in radio direction finding is a crucial technology in the field of communications. It identifies anomalous data points by learning patterns in historical direction finding data, helping to identify and eliminate anomalous data that negatively impacts signal direction finding accuracy, thereby improving the accuracy of signal direction finding. Simultaneously, in counter-drone systems, accurate antenna direction finding can improve target positioning accuracy, reduce false alarm rates, enhance system robustness, and effectively monitor the activities and intentions of unauthorized drones. Existing anomaly detection methods can generally be divided into two categories: statistical anomaly detection methods and deep learning-based anomaly detection network methods. Statistical anomaly detection methods assume that azimuth data conforms to a specific statistical distribution and detect anomalies at the current time point based on the distribution of past data. However, they may perform poorly when facing atypical distributions, resulting in poor anomaly detection performance, especially in complex electromagnetic environments.

[0003] Currently, research on anomaly detection networks based on deep learning mainly focuses on time series algorithms. Existing studies have compared the outlier detection performance of Gaussian statistical models, local outlier factors, support vector machines, isolated forest algorithms, and long short-term memory neural networks. Experiments have shown that deep learning-based methods achieve higher accuracy than traditional statistical and machine learning outlier detection methods. Although long short-term memory neural networks excel at capturing temporal dependencies, they have limitations in handling spatial relationships in data, which may restrict their performance on certain datasets. Other studies employ deep autoencoders to capture the spatial features of multivariate time series data and combine them with multi-scale sparse transformer networks to extract temporal features. Simultaneously, experiments have demonstrated that outlier detection networks with spatiotemporal feature extraction modules can achieve better performance than long short-term memory neural networks. Radio direction finding technology is affected by multipath effects, electromagnetic interference, weather condition changes, and spectral aliasing, negatively impacting the accuracy and performance of direction finding. Patent CN116451118B discloses an outlier detection method based on deep learning for radar and electro-optical direction finding data. This method utilizes deconvolutional coding to fuse the temporal and spatial features of multi-sensor point data and reconstruct the feature matrix. It then calculates a loss function with the original feature matrix to detect outliers when sensors detect targets. This addresses, to some extent, the problems caused by labeling large amounts of multimodal data. However, this unsupervised learning method has several limitations. First, deep learning models are often considered "black boxes," making it difficult to explain why the model labels a data point as an outlier. Without labels, some outliers might be considered normal. Second, unsupervised models typically require large amounts of data for training, and hyperparameters are difficult to determine, potentially affecting model performance.

[0004] These deep learning-based methods all assume the availability of large amounts of data for network training, which is unrealistic in practical scenarios. This is because radio direction finding data is typically sparse, while deep learning models require vast amounts of data for training, which may be difficult to obtain in real-world environments. Furthermore, many existing algorithms are time-series based, which are insufficient in capturing spatial information and structure, making it difficult to effectively represent complex spatial correlations and image structural features. They may also perform poorly when processing data with complex spatial distributions and structures. Therefore, there is an urgent need to develop a new deep learning-based method for radio direction finding anomaly detection, taking into account multiple perspectives such as spatial and temporal correlations and combining them with the characteristics of radio direction finding data. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a radio direction finding anomaly detection method based on image segmentation. It models anomaly detection as a vertex segmentation problem, utilizes deep neural networks to learn high-level features of the data, and simultaneously mines azimuth context information. This not only endows the network with interpretability but also improves the performance of radio direction finding anomaly detection through spatial correlation and local fine-grained feature capture. It overcomes the limitations of handling complex noise and possesses advantages such as strong spatial correlation, high interpretability, and strong transferability. It can be applied to practical radio direction finding scenarios, improving the accuracy of radio direction finding and solving problems such as low accuracy and poor reliability caused by signal interference and multipath effects in radio direction finding. It overcomes the shortcomings of existing anomaly detection technologies, such as poor performance and poor transferability in practical communication scenarios, which cannot meet the requirements of different application scenarios, and the lack of interpretability in traditional deep learning anomaly detection methods.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0007] A radio direction-finding anomaly detection method based on image segmentation, comprising:

[0008] Step 1: The array antenna receives signals from the drone to achieve real radio data acquisition;

[0009] Step 2: Construct a multi-signal classification spatial spectrum estimation function based on the UAV signal, and extract the azimuth angle of the UAV signal corresponding to the peak value of the function;

[0010] Step 3: By using a sliding window, the azimuth data slices are reassembled into a two-dimensional matrix set. At the same time, the real azimuth of the UAV is calculated based on the real latitude and longitude data of the UAV. The training dataset is constructed using the two-dimensional image structure matrix and the real azimuth of the UAV.

[0011] Step 4: Construct a radio direction finding anomaly detection network based on an encoder and decoder structure;

[0012] Step 5: Determine the optimization objective of the radio direction finding anomaly detection network and adjust the loss function used for the network parameters;

[0013] Step 6: Based on the optimization objective and loss function, train the radio direction finding anomaly detection network using the training dataset;

[0014] Step 7: Input the test data into the trained radio direction finding anomaly detection network and output pixel-level azimuth prediction results.

[0015] To optimize the above technical solution, the specific measures also include:

[0016] The UAV signal received by the array antenna in step 1 above is:

[0017] X t =M t S t +ω t

[0018] in, This represents the array manifold matrix received at second t, and the steering vector. Let θ1 represent the direction of arrival of the UAV signals received by the N array elements, θ1 represent the angle of arrival of the UAV signals, λ and d represent the signal wavelength and the radius of the circular array, respectively, and j is the imaginary unit. T It is the transpose operation of a vector; S t Represents the signal source vector; ω t This indicates that the mean is 0 and the variance is σ. 2 Additive white Gaussian noise.

[0019] Step 2 above includes:

[0020] (2a) Calculate the signal subspace and noise subspace based on the UAV signal:

[0021] First, calculate the covariance matrix:

[0022]

[0023] Among them, R x This represents the UAV signal X received by the array antenna in the t-th second. t The covariance matrix; S represents the signal source vector. t The covariance matrix; I represents the identity matrix, H is the conjugate transpose operation, and E{·} is the expectation operation;

[0024] Then, for R x Eigenvalue decomposition yields eigenvalues ​​and eigenvectors, which in turn create the signal subspace and noise subspace.

[0025] rank(R x ) = 1, thus obtaining the signal subspace Us = (q0) and the noise subspace U N =(q1,q2,…,q N-1 );

[0026] Among them, due to rank(R) x If β = 1, it has only one non-zero eigenvalue, and all other eigenvalues ​​are zero. i Let q0 represent the i-th eigenvalue, where q0 is the eigenvector corresponding to the non-zero β0, and the others q1, q2, ..., q... represent the eigenvalues. N-1 It is the eigenvector corresponding to the zero eigenvalue.

[0027] (2b) Construct a multi-signal classification spatial spectrum estimation function based on the noise subspace:

[0028]

[0029] P MUSIC The value of θ1 corresponding to the maximum value of (θ1) is the azimuth angle of the UAV signal.

[0030] Step 3 above includes:

[0031] The first step is to initialize the sliding window parameters c and l = max{|A1|,|A2|,…,|A...}. t |};

[0032] Where c represents the number of samples contained in the sliding window, l represents the length of each row of azimuth data in the sliding window, and A t Let represent the set of azimuth angle data of the UAV received by the array antenna in the t-th second, |·| represent the number of elements in the vector, and max represents the maximum value operation;

[0033] The second step is to use a sliding window. With a step size of 1, the azimuth data set is sliced ​​and reassembled into a two-dimensional matrix set, as shown in the following formula:

[0034] W = [W0, W1, ..., W r-c ]

[0035] W i =A[i:i+c,0:l],i=0,1,…,rc

[0036] Where W represents the set of sliding window slice data, i.e., the set of two-dimensional matrices; W i W represents the data after the i-th sliding window slice. i It is a c×l two-dimensional matrix whose elements belong to the set of real numbers. A represents the azimuth angle data collected by the antenna within a continuous time period. Each row represents a moment in the continuous time period, storing the azimuth angle data collected by the antenna at that moment. t A azimuth angle, if |A t If |≤l, pad with zeros to keep the azimuth length consistent at l; r=|A| represents the number of samples acquired by the antenna, and c represents the number of samples contained in the sliding window when slicing;

[0037] The third step is to obtain the true azimuth angle of the UAV relative to due north by using the latitude and longitude of the array antenna and the real-time latitude and longitude of the UAV.

[0038] The fourth step is to add labels to the data in the two-dimensional matrix set according to the actual azimuth angle of the UAV relative to due north in a counterclockwise direction, to form a training dataset. In the training dataset, the value closest to the actual azimuth angle in each timestamp is marked as 1 in the matrix, while other azimuth angles are marked as 0.

[0039] The third step mentioned above is as follows:

[0040] Based on the latitude and longitude (lon1, lat1) of the array antenna and the real-time latitude and longitude (lon2) of the UAV at second t. t ,lat2 t To obtain the true azimuth angle θ of the UAV relative to true north at second t. t ':

[0041] θ t =atan2(sin(lon2) t -lon1),cos(lat1)*tan(lat2 t )-sin(lat1)*cos(lon2 t -lon1)).

[0042] The input to the radio direction finding anomaly detection network described in step 4 above is a set of two-dimensional matrices composed of azimuth angles. The encoder encodes the input two-dimensional matrix set, and the encoding process is as follows:

[0043] F1 = (maxpool) 22 (W*K1))*K1

[0044] Where F1 represents the high-level features output by the encoder; * represents the convolution operation; K1 represents a convolution block, consisting of two sets of 3×3 convolution kernels, each set of kernels is followed by a normalization layer (BN) and a ReLU activation function f(x) = max(0,x), and maxpooling. 22 This represents 2×2 max pooling.

[0045] The decoding process of the above decoder is as follows:

[0046] F2=sigmoid(K4*(K3*(concat((K2*F1),x1))))

[0047] Where F2 is the pixel-level azimuth prediction result output by the encoder; x1 represents the feature map output by W in the encoder after passing through the K1 convolutional block; K2 represents an upsampled convolutional block, including an upsampling operation with a scaling factor of 2, followed by a 3×3 convolutional kernel, BN, and ReLU activation function; concat represents the concatenation operation; K3 is a convolutional block, and K4 is a 1×1 convolutional kernel; the sigmoid function is sigmoid(x) = 1 / (1+e^(-x)).(-x) ).

[0048] The optimization objective of the radio direction finding anomaly detection network described in step 5 above is:

[0049]

[0050] Where, θ f Let |W| represent the parameters of the spectral outlier detection network based on image segmentation, and |W| represent the number of two-dimensional data slices. and W i The actual label and the predicted label output by the network for the slice in row h and column k.

[0051] The present invention has the following beneficial effects:

[0052] First, this invention proposes a new, general, and interpretable outlier detection method that utilizes image segmentation to detect outliers in antenna azimuth data. Compared to traditional deep learning frameworks, this method overcomes the black-box effect of networks and increases the interpretability of radio direction finding anomaly detection networks.

[0053] Second, the image segmentation method of this invention utilizes contextual information within a window for anomaly detection, eliminating the need for feature selection and design. It can capture local spatial relationships and structural information, taking into account the temporal and spatial correlation of the azimuth angle data, and overcoming the limitations of processing complex noise, thereby improving the accuracy and reliability of antenna azimuth angle anomaly detection.

[0054] Third, compared with traditional statistical and machine learning methods, this invention utilizes the convolution operation in deep neural networks to overcome the limitations of handling nonlinear noise, and realizes the detection of outlier values ​​in the azimuth angle under complex electromagnetic environments, which can be applied to practical radio target detection systems. Attached Figure Description

[0055] Figure 1 This is a flowchart of the radio direction finding anomaly detection method based on image segmentation according to the present invention;

[0056] Figure 2 This paper compares the detection method of the present invention with traditional statistical outlier detection methods under different tolerance angles and intersection-to-union ratios (IOU). Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] Although the steps in this invention are arranged by reference numerals, this is not intended to limit the order of the steps. Unless the order of the steps is explicitly stated or the execution of a step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" as used herein refers to and covers any and all possible combinations of one or more of the associated listed items.

[0059] like Figure 1 As shown, the present invention provides a radio direction finding anomaly detection method based on image segmentation, comprising:

[0060] Step 1: The array antenna receives signals from the drone to achieve real radio data acquisition;

[0061] This invention constructs a UAV monitoring platform composed of an N-element uniform circular array. Through a signal reception model, the receiving end of the array antenna can capture the UAV's downlink image transmission signal. Assuming the existence of a desired far-field narrowband UAV signal, the signal received by the array antenna is:

[0062] X t =M t S t +ω t

[0063] in, This represents the array manifold matrix received at second t, and the steering vector. Let θ1 represent the direction of arrival of the UAV signals received by the N array elements, θ1 represent the angle of arrival of the UAV signals, λ and d represent the signal wavelength and the radius of the circular array, respectively, and j is the imaginary unit. T It is the transpose operation of a vector; S t Represents the signal source vector; ω t This indicates that the mean is 0 and the variance is σ. 2 Additive white Gaussian noise.

[0064] In an embodiment of the invention, a five-element array antenna is used to receive signals from the UAV. When the UAV appears within the antenna's detection area, the antenna captures the UAV's downlink image transmission link signal. Typically, UAVs communicate using the 2.4GHz and 5.8GHz frequency bands. This invention assumes that the frequency band of the UAV's image transmission link is fixed and does not employ frequency hopping communication technology.

[0065] Step 2: Construct a multi-signal classification spatial spectrum estimation function based on the UAV signal, and extract the azimuth angle of the UAV signal corresponding to the peak value of the function to realize the direction finding azimuth angle extraction.

[0066] Based on the UAV image transmission link signal collected in step 1, the azimuth angle of the UAV signal corresponding to the peak value is found by constructing a MUSIC spatial spectrum estimation function:

[0067] (2a) Calculate the signal subspace and noise subspace:

[0068]

[0069] Among them, R x This represents the UAV signal X received by the array antenna in the t-th second. t The covariance matrix; S represents the signal source vector. t The covariance matrix is: I represents the identity matrix, H is the conjugate transpose operation, and E{·} is the expectation operation.

[0070] Then, for R x Eigenvalue decomposition yields eigenvalues ​​and eigenvectors.

[0071] Since there is only one drone signal source, the signal and noise are uncorrelated, therefore rank{R x} = 1, the signal subspace is represented as Us = (q0), and the noise subspace is represented as U N =(q1,q2,…,q N-1 ), β i Let q0 represent the i-th eigenvalue, where q0 is the eigenvector corresponding to the non-zero β0, and the others q1, q2, ..., q... represent the eigenvalues. N-1 It is the eigenvector corresponding to the zero eigenvalue.

[0072] (2b) Constructing the spatial spectrum estimation function

[0073]

[0074] P MUSIC The value of θ1 corresponding to the maximum value of (θ1) is the azimuth angle of the UAV signal.

[0075] Step 3: Initialize the sliding window parameters. Through the sliding window, the azimuth dataset is sliced ​​and reconstructed into a two-dimensional image structure. At the same time, the real azimuth of the UAV is calculated based on the real latitude and longitude data of the UAV. The training dataset is constructed with the two-dimensional matrix set and the real azimuth of the UAV.

[0076] First, the window size, sliding step, and filling method of the sliding window need to be initialized. The azimuth dataset acquired by the antenna is sliced ​​into several matrices and reassembled into the structure of the two-dimensional image set, as follows:

[0077] The first step is to initialize the sliding window parameters c and l = max{|A1|,|A2|,…,|A...}.t |}.

[0078] Where c represents the number of samples contained in the sliding window, l represents the length of each row of azimuth data in the sliding window, and A t Let represent the set of azimuth angles received by the array antenna in the t-th second, |·| represent the number of elements in the vector, max represents the maximum value operation, and the sliding step size is initialized to 1;

[0079] The second step is to use a sliding window. With a step size of 1, the azimuth data set is sliced ​​and reassembled into a two-dimensional matrix set, as shown in the following formula:

[0080] W = [W0, W1, ..., W r-c ]

[0081] W i =A[i:i+c,0:l],i=0,1,…,rc

[0082] Where A represents the azimuth data acquired by the antenna within a continuous time period, and each row represents a moment in that continuous time (in seconds), storing the azimuth data acquired by the antenna at that moment. t A azimuth angle. If |A t If |≤l, pad with zeros to ensure the azimuth length remains constant at l, r=|A| represents the number of samples collected by the antenna, W i Let W represent the data after the i-th sliding window slice, and let W represent the set of data sliced ​​by the sliding window. The step size is set to 1, which enables continuous sliding between windows. Each data point has the opportunity to be included in multiple windows, thus capturing the local features of the spectral data more accurately. In addition, using a sliding window to slice the azimuth data can obtain a large number of training samples, increasing the diversity and richness of the data.

[0083] Then, the true azimuth angle of the UAV is calculated based on the real-time latitude and longitude data of the UAV. A training dataset is constructed using a set of two-dimensional matrices and the true azimuth angle of the UAV, as follows:

[0084] The third step is to record the real-time latitude and longitude data of the UAV and calculate the true azimuth angle of the UAV relative to the array antenna. This is done based on the latitude and longitude of the array antenna (lon1, lat1) and the real-time latitude and longitude of the UAV at second t (lon2). t ,lat2 t To obtain the true azimuth angle θ of the UAV relative to true north at second t. t ';

[0085] θ t =atan2(sin(lon2) t-lon1),cos(lat1)*tan(lat2 t )-sin(lat1)*cos(lon2 t -lon1))

[0086] The fourth step is to add labels to the data in the training set based on the actual azimuth angle of the drone. In the training set, the value closest to the actual azimuth angle within each timestamp is marked as 1 in the matrix, while other azimuth angles are marked as 0.

[0087] Machine learning-based outlier detection requires large amounts of sample data and presents challenges in feature engineering, model selection, and hyperparameter tuning. Furthermore, its black-box architecture makes the network lack interpretability. In contrast, image segmentation-based methods utilize contextual information within a window for outlier detection, eliminating the need for feature selection and design, and capturing local spatial relationships and structural information. Simultaneously, image segmentation methods can visualize and interpret detected outliers, exhibiting advantages such as good adaptability to unknown data distributions, strong interpretability, and robustness.

[0088] Step 4: Construct a radio direction finding anomaly detection network based on encoder and decoder structure.

[0089] First, the input is a set of two-dimensional matrices W = [W0, W1, ..., W...] composed of azimuth angles sliced ​​through a sliding window. r-c ], where r represents the number of samples collected by the antenna; W i This represents the data after the i-th sliding window slice, containing c rows of sample data, with the number of columns l = max{|A1|,|A2|,…,|A...} t |} is the maximum number of azimuth angles collected at each time point.

[0090] The encoder gradually reduces the spatial dimensionality of the input image through convolution, pooling, and non-linear activation function operations, compressing and abstracting noise and detail information in the image, capturing high-level features while preserving the main structural information of the image. The encoding process can be represented as follows:

[0091] F1 = (maxpool) 22 (T*K1))*K1

[0092] Where * denotes a convolution operation, K1 represents a convolution block consisting of two sets of 3×3 convolution kernels, each followed by a normalization layer (BN) and a ReLU activation function f(x) = max(0,x), and maxpooling. 22 This represents 2×2 max pooling.

[0093] Next, the decoder gradually enlarges the feature maps in the encoder through deconvolution and upsampling operations, restoring them to the original image size. Simultaneously, skip connections are used to fuse low-level features from the encoder with high-level features from the decoder to obtain more accurate prediction results. The decoding process can be represented as follows:

[0094] F2=sigmoid(K4*(K3*(concat((K2*F1),x1))))

[0095] Here, F1 represents the high-level features output by the encoder, x1 represents the feature map output by the encoder's W after passing through the K1 convolutional block, K2 represents the upsampling convolutional block, which includes an upsampling operation with a scaling factor of 2, followed by a 3×3 convolutional kernel, BN, and the ReLU activation function. concat represents the concatenation operation, which can reduce the loss of fine-grained information caused by downsampling. The output of K2 is concatenated with the x1 feature map in the encoder along the first dimension. Then, the merged feature map is passed through the K3 convolutional block and the K4 1×1 convolutional kernel to restore the original image size and number of channels. Finally, sigmoid(x) = 1 / (1+e (-x) The function generates pixel-level azimuth prediction results.

[0096] Step 5: Define the optimization objective of the radio direction finding anomaly detection network, adjust the network parameters to minimize the loss function, and improve the performance of anomaly detection.

[0097] In the problem of outlier detection based on image segmentation, optimization requires defining a metric function to measure the difference between the segmentation result and the true label. The optimization objective can be expressed by the following formula:

[0098]

[0099] Where, θ f Let |W| represent the parameters of the spectral outlier detection network based on image segmentation, and |W| represent the number of two-dimensional data slices. and W i The actual label and the predicted label output by the network for the slice in row h and column k.

[0100] The metric uses the cross-entropy loss function. By accumulating the minimum loss function, the optimization goal can be achieved, namely, minimizing the difference between the segmentation result and the true label. The model can more accurately detect and locate outliers, thus improving the accuracy of the outlier detection algorithm.

[0101] By solving the above optimization problem, the optimal segmentation probability matrix can be obtained. Based on this probability matrix, the collected azimuth angles are segmented into two categories: outliers and UAV azimuth angles, thus achieving the task of outlier detection in spectral azimuth angles. It is important to note that in a new scenario, the model can directly perform outlier detection on the azimuth angle data collected by the antenna. Simultaneously, it can label some newly collected data to further fine-tune the model's parameters, enabling more accurate identification and location of outliers and improving its performance in new scenarios. This gives the model good transfer and adaptability, allowing it to adapt to outlier detection tasks in different scenarios.

[0102] Step 6: Train the network model using the training data.

[0103] The first step is to randomly initialize the trainable parameters of the network, setting the initial training iteration count to 1 epoch, the maximum number of iterations to 300, and the learning rate to 0.001. The Adam optimization algorithm is used as the network training optimizer. If the model needs to be applied to a completely new data acquisition scenario, the network can be initialized using pre-trained model parameters. This can be achieved by loading a pre-trained model and then applying the parameters to the current network.

[0104] The second step involves inputting the training data into the network in batches. The batch size can be adjusted, and the training error of each batch is backpropagated to optimize the network parameters. One epoch is completed when all batches of training data have been backpropagated.

[0105] The third step is to determine whether the network training is complete: If the network training is complete, proceed to step 7; otherwise, increment the training iteration count by one and continue training the network.

[0106] Step 7: Input the test set data into the network and output a pixel-level probability matrix, dividing the data into two categories: background noise and target azimuth.

[0107] The effects of the present invention will be further explained below with reference to simulation experiments.

[0108] 1. Simulation conditions and parameter settings:

[0109] The simulation experiments of this invention were conducted on a simulation platform using Python 3.8 and PyTorch 2.1.2. The computer CPU was an Intel Core i7, equipped with an NVIDIA GeForce RTX 3060 dedicated graphics card. The azimuth angle of the drone was selected based on the frequency band. The sampling frequency of the array antenna was 2048 Hz, the sliding window value was set to 39, the step size was 1, the maximum number of iterations for network training was 300, the learning rate was 0.001, the Adam optimization algorithm was selected as the network training optimizer, and the batch size was 32.

[0110] Traditional image segmentation evaluation metrics require that the boundaries of detection boxes or segmentation results perfectly match the boundaries of the ground truth labels. This strict matching requirement may lead to significant errors in the model's selection of angles close to the true azimuth. To increase the robustness and fault tolerance of the model in outlier detection tasks, this invention uses the Intersection over Union (IOU) metric with tolerance angles. Considering that the relative angular tolerance between bounding boxes is within 3°, the ratio of the intersection area to the union area of ​​the bounding boxes of two segmented regions is calculated:

[0111]

[0112] In this model, TP represents the number of positive class samples correctly predicted as positive, FP represents the number of negative class samples incorrectly predicted as positive, and FN represents the number of positive class samples incorrectly predicted as negative. Typically, the IoU value ranges from 0 to 1; a value closer to 1 indicates a higher degree of overlap between the predicted result and the true label, suggesting more accurate class segmentation.

[0113] 2. Simulation content:

[0114] Figure 2 This paper compares the proposed image segmentation-based outlier detection algorithm with traditional statistical outlier detection methods under different tolerance angles using the Intersection over Union (IOU) ratio. Since radio direction finding systems require high target accuracy, the allowable error angle range in the IOU is from 1° to 3°. The comparison shows that the denoising accuracy of the proposed algorithm is significantly higher than other existing methods. Overall, the overlap between the acquired radio azimuth angle and the true azimuth angle increases with the increase of the allowable error angle. With a tolerance angle of 3°, the proposed algorithm achieves an accuracy of approximately 97.4% in the same acquisition scenario as the training set, and an accuracy of 98.86% when transferring the model to different scenarios. This indicates that the image segmentation-based outlier detection algorithm can control most of the acquired azimuth angle errors within 3°, which exceeds the accuracy of statistical maximum outlier detection methods by approximately 50% and statistical mean outlier detection methods by approximately 10%. When the tolerance angle is less than 3°, it is evident that the image segmentation-based outlier detection algorithm proposed in this invention is more competitive and robust under stricter matching requirements. Furthermore, comparing the accuracy of the three tolerance angles in different scenarios, the outlier detection algorithm proposed in this invention, compared to statistical outlier detection methods, demonstrates strong generalization ability in unknown and complex electromagnetic environments. It can transfer knowledge learned from past models to new environments while maintaining high accuracy.

[0115] Based on the above simulation results and analysis, the radio direction finding anomaly detection method based on image segmentation proposed in this invention can achieve higher anomaly removal accuracy than existing methods, and has strong transferability, can adapt to new radio azimuth angle acquisition environments, and has strong network interpretability, which enables this invention to be better applied in actual communication scenarios.

[0116] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0117] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for detecting radio direction-finding anomalies based on image segmentation, characterized in that, include: Step 1: The array antenna receives the UAV signal to achieve real radio data acquisition; the UAV signal received by the array antenna is: X t =M t S t +oh t in, This represents the array manifold matrix received at second t, and the steering vector. Let θ1 represent the direction of arrival of the UAV signals received by the N array elements, θ1 represent the angle of arrival of the UAV signals, λ and d represent the signal wavelength and the radius of the circular array, respectively, and j is the imaginary unit. T It is the transpose operation of a vector; S t Represents the signal source vector; ω t This indicates that the mean is 0 and the variance is σ. 2 Additive white Gaussian noise; Step 2: Construct a multi-signal classification spatial spectrum estimation function based on the UAV signal, and extract the azimuth angle of the UAV signal corresponding to the peak value of the function, including: (2a) Calculate the signal subspace and noise subspace based on the UAV signal: First, calculate the covariance matrix: Among them, R x This represents the UAV signal X received by the array antenna in the t-th second. t The covariance matrix; S represents the signal source vector. t The covariance matrix; I represents the identity matrix, H is the conjugate transpose operation, and E{·} is the expectation operation; Then, for R x Eigenvalue decomposition yields eigenvalues ​​and eigenvectors, which in turn create the signal subspace and noise subspace. rank(R x ) = 1, thus obtaining the signal subspace Us = (q0) and the noise subspace U N =(q 1, q 2, …,q N-1 ); Where, β i Let q0 represent the i-th eigenvalue, where q0 is the eigenvector corresponding to the non-zero β0, and the others q... 1, q 2, …,q N-1 It is the eigenvector corresponding to the zero eigenvalue; (2b) Construct a multi-signal classification spatial spectrum estimation function based on the noise subspace: P MUSIC The value of θ1 corresponding to the maximum value of (θ1) is the azimuth angle of the UAV signal; Step 3: Using a sliding window, the azimuth data slices are reassembled into a two-dimensional matrix set. Simultaneously, the UAV's true azimuth is calculated based on its real-time latitude and longitude data. A training dataset is constructed using the two-dimensional matrix set and the UAV's true azimuth, including: The first step is to initialize the sliding window parameters c and l = max{|A1|,|A2|,…,|A...}. t }; Where c represents the number of samples included in the sliding window during slicing, l represents the length of each row of azimuth data in the sliding window, and A t Let represent the set of azimuth angle data of the UAV received by the array antenna in the t-th second, |·| represent the number of elements in the vector, and max represents the maximum value operation; The second step is to use a sliding window. With a step size of 1, the azimuth data set is sliced ​​and reassembled into a two-dimensional matrix set, as shown in the following formula: W=[W0,W1,…,W r-c ] W i =A[i:i+c,0:l],i=0,1,…,r-c Where W represents the set of sliding window slice data, i.e., the set of two-dimensional matrices; W i W represents the data after the i-th sliding window slice. i It is a c×l two-dimensional matrix whose elements belong to the set of real numbers. A represents the azimuth angle data collected by the antenna within a continuous time period. Each row represents a moment in the continuous time period, storing the azimuth angle data collected by the antenna at that moment. t A azimuth angle, if |A t If |≤l, pad with zeros to keep the azimuth length consistent at l; r=|A| represents the number of samples collected by the antenna; The third step is to obtain the true azimuth angle of the UAV relative to due north by using the latitude and longitude of the array antenna and the real-time latitude and longitude of the UAV. The fourth step is to add labels to the data in the two-dimensional matrix set according to the actual azimuth angle of the UAV relative to due north in a counterclockwise direction, to form a training dataset. In the training dataset, the value closest to the actual azimuth angle in each timestamp is marked as 1 in the matrix, while other azimuth angles are marked as 0. Step 4: Construct a radio direction finding anomaly detection network based on an encoder and decoder structure; Step 5: Determine the optimization objective of the radio direction finding anomaly detection network and adjust the loss function used for the network parameters; Step 6: Based on the optimization objective and loss function, train the radio direction finding anomaly detection network using the training dataset; Step 7: Input the test data into the trained radio direction finding anomaly detection network and output pixel-level azimuth prediction results.

2. The radio direction finding anomaly detection method based on image segmentation according to claim 1, characterized in that, The third step is as follows: Based on the latitude and longitude (lon1, lat1) of the array antenna and the real-time latitude and longitude (lon2) of the UAV at second t. t ,lat2 t To obtain the true azimuth angle θ of the UAV relative to true north at second t. t ': θ' t =time2(sin(long2 t -long1),cos(lat1)*time(lat2 t )-sin(lat1)*cos(long2 t -lon1))。 3. The radio direction finding anomaly detection method based on image segmentation according to claim 1, characterized in that, The input to the radio direction finding anomaly detection network described in step 4 is a set of two-dimensional matrices composed of azimuth angles. The encoder encodes the input two-dimensional matrix set, and the encoding process is as follows: F1=(maxpool 22 (W*K1))*K1 Where F1 represents the high-level features output by the encoder; W is the set of sliding window slice data, i.e., the set of two-dimensional matrices; * represents the convolution operation; K1 represents the convolution block, which consists of two sets of 3×3 convolution kernels. Each set of convolution kernels is followed by a normalization layer BN and a ReLU activation function f(x) = max(0,x), and maxpooling. 22 This represents 2×2 max pooling.

4. The radio direction finding anomaly detection method based on image segmentation according to claim 1, characterized in that, The decoding process of the decoder is as follows: F2=sigmoid(K4*(K3*(concat((K2*F1),x1)))) Where F2 is the pixel-level azimuth prediction result output by the encoder; x1 represents the feature map output by W in the encoder after passing through the K1 convolutional block; K2 represents an upsampled convolutional block, including an upsampling operation with a scaling factor of 2, followed by a 3×3 convolutional kernel, BN, and ReLU activation function; concat represents the concatenation operation; K3 is a convolutional block, and K4 is a 1×1 convolutional kernel; the sigmoid function is sigmoid(x) = 1 / (1+e^(-x)). (-x) ).

5. The radio direction finding anomaly detection method based on image segmentation according to claim 1, characterized in that, The optimization objective of the radio direction finding anomaly detection network described in step 5 is: Where, θ f Let |W| represent the parameters of the spectral outlier detection network based on image segmentation, and |W| represent the number of two-dimensional data slices. and W i The true label and the predicted label output by the network for the slice in the h-th row and k-th column, W. i Let represent the data after the i-th sliding window slice, c represent the number of samples included in the sliding window during slicing, and l represent the length of each row of azimuth data in the sliding window.

Images