A radar abnormal data intelligent detection and identification method and system

Abstract

This invention discloses an intelligent detection and identification method and system for radar anomaly data, relating to the field of radio technology. The method includes: acquiring multi-channel raw echo data from a phased array radar system in real time; extracting multi-dimensional feature vectors from the raw echo data, wherein the extracted multi-dimensional feature vectors include time-domain features, frequency-domain features, and spatial-domain features, with the spatial-domain features including the beam pattern features after beamforming; constructing an anomaly detection model, inputting the multi-dimensional feature vectors into the anomaly detection model, and outputting anomaly detection results; when the detection result is an anomaly, identifying the anomaly type; generating corresponding adjustment instructions based on the anomaly type identification results, uploading the anomaly information, and subsequently issuing and executing the system adjustment instructions. This invention solves the technical problems of existing technologies that rely on a single parameter, are susceptible to external environmental interference, and cannot identify the specific type of anomaly, leading to inaccurate anomaly identification and a high false alarm rate in anomaly detection.

Description

Technical Field

[0001] This invention relates to the field of radio technology, and more specifically to a method and system for intelligent detection and identification of radar anomaly data. Background Technology

[0002] As a core component of modern detection systems, phased array radar's performance and reliability are directly related to critical tasks such as air defense, missile defense, and battlefield surveillance. Under complex electromagnetic environments and high-load operation, radio wave echo data is susceptible to various abnormal factors.

[0003] Currently, radar anomaly detection mainly relies on a single parameter. However, due to the susceptibility of detection to external environmental interference and the inability to identify the specific type of anomaly, it suffers from inaccurate anomaly identification and a high false alarm rate. Summary of the Invention

[0004] This application provides a method and system for intelligent detection and identification of radar anomaly data, which addresses the technical problems of existing technologies that rely on a single parameter, are easily affected by external environmental interference, cannot identify the specific type of anomaly, and thus result in inaccurate anomaly identification and a high false alarm rate.

[0005] In view of the above problems, this application provides a method and system for intelligent detection and identification of radar anomaly data.

[0006] Firstly, this application provides a method for intelligent detection and identification of radar anomaly data, the method comprising: Real-time acquisition of multi-channel raw echo data from phased array radar systems; Extracting multidimensional feature vectors from raw echo data, including time-domain features, frequency-domain features, and spatial-domain features, wherein the spatial-domain features include the beam pattern features after beamforming; Construct an anomaly detection model, input the multidimensional feature vector into the anomaly detection model, and output the anomaly detection result; When the detection result is abnormal, the type of abnormality is identified; Based on the anomaly type identification results, corresponding system adjustment instructions are generated, the anomaly information is uploaded, and then the system adjustment instructions are issued and executed.

[0007] Secondly, the present invention provides an intelligent detection and identification system for radar anomaly data, comprising: The data acquisition module is used to acquire multi-channel raw echo data from the phased array radar system in real time; The feature extraction module is used to extract multi-dimensional feature vectors from the raw echo data. The extracted multi-dimensional feature vectors include time-domain features, frequency-domain features, and spatial-domain features. The spatial-domain features include the beam pattern features after beamforming. Anomaly detection module is used to construct anomaly detection model, input the multidimensional feature vector into the anomaly detection model, and output anomaly detection results; The anomaly classification module identifies the anomaly type when the detection result is abnormal. The instruction execution module generates corresponding system adjustment instructions based on the anomaly type identification results, uploads the anomaly information, and then issues and executes the system adjustment instructions.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application firstly ensures the comprehensiveness of observation data for anomaly detection processing by acquiring multi-channel raw echo data from a phased array radar system in real time, preserving complete time-domain, frequency-domain, and spatial-domain data to provide multi-dimensional information input for subsequent processing. Secondly, it extracts multi-dimensional feature vectors containing time-domain, frequency-domain, and beamforming pattern features from the raw echo data, constructing a multi-dimensional fusion feature space in the time, frequency, and spatial domains. This significantly enhances the ability to identify various anomaly types, avoiding single-feature or single-dimensional analysis. Furthermore, it updates the anomaly detection model through reinforcement learning algorithms for feedback adjustment. Thirdly, it constructs an anomaly detection model and inputs the multi-dimensional feature vectors, outputting detection results. A multi-branch deep learning network is constructed using machine learning algorithms to acquire nonlinear fitting and pattern recognition capabilities, enabling the detection of complex anomaly patterns and improving detection accuracy. Simultaneously, when the detection result is an anomaly, the anomaly type is identified, and the specific anomaly type and anomaly confidence level are output, providing a decision-making basis for subsequent precise countermeasures. Finally, based on the anomaly type identification results, corresponding system adjustment instructions are generated, and information is uploaded and instructions are executed. This enables the radar system to make adjustments for different situations such as hardware failures, environmental interference, and abnormal target characteristics. Complex anomaly information is also coordinated with the command center, ultimately improving the reliability of the radar system and its target identification capabilities. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a radar anomaly data intelligent detection and identification method according to this application; Figure 2 This is a schematic diagram of the structure of a radar anomaly data intelligent detection and identification system according to this application.

[0010] In the attached diagram, the components represented by each number are as follows: Data acquisition module 11, feature extraction module 12, anomaly detection module 13, anomaly classification module 14, instruction execution module 15. Detailed Implementation

[0011] This application provides an intelligent detection and identification method for radar anomaly data, which specifically solves the technical problems of existing technologies that rely on a single parameter, are easily affected by external environmental interference, cannot identify the specific type of anomaly, and thus result in inaccurate anomaly identification and a high false alarm rate.

[0012] The present invention will now be described in detail with reference to the accompanying drawings.

[0013] Example 1, as Figure 1 As shown, this application provides an intelligent detection and identification method for radar anomaly data, the method comprising: S10: Real-time acquisition of multi-channel raw echo data from the phased array radar system; In this embodiment, a phased array radar is a radar that changes the beam direction by controlling the phase of each radiating element in the array antenna; a phased array radar system is the management or operating system of the phased array radar, which contains the operation or running data of the phased array radar; raw echo data is the echo signal received after the radar transmits a signal without or only after preliminary digitization processing, and usually contains complex forms of information such as target reflection, clutter, interference and noise.

[0014] Specifically, firstly, within the area where the phased array radar is located, raw echo data from all receiving channels of the phased array radar system over a certain period of time, such as the past two years, is collected in real time and synchronously to obtain a large amount of multi-channel data in the time domain, frequency domain, and spatial domain, providing a sufficient data foundation for subsequent anomaly identification.

[0015] In this embodiment, by synchronously collecting all receiving channels in real time, the raw echo data of multiple channels is obtained, and the phase and amplitude relationships between each channel are obtained, ensuring that abnormal conditions can be captured from multiple domains such as time domain, frequency domain, and spatial domain, providing rich detection data for subsequent feature extraction.

[0016] S20: Extract multi-dimensional feature vectors from the raw echo data, wherein the extracted multi-dimensional feature vectors include time-domain features, frequency-domain features and spatial-domain features, and the spatial-domain features include the beam pattern features after beamforming. In this embodiment, time-domain features are the attribute features of phased array radar in the time dimension; frequency-domain features are the attribute features of phased array radar in the frequency dimension; spatial features are the attribute features of phased array radar in the spatial dimension; beamforming is achieved by weighting the received signals of each channel so that the array antenna forms a high-gain receiving beam in a specific direction while suppressing interference in other directions.

[0017] Specifically, features are extracted in parallel from all acquired raw echo data across three different physical dimensions: time domain, frequency domain, and spatial domain. Time domain features are extracted from the signal waveform itself; frequency domain features are extracted from the signal's spectral structure; and multi-channel data are synthesized into a spatially directional beam using beamforming processing, from which spatial domain features are extracted. Finally, the feature subsets extracted from different dimensions are normalized and concatenated to form a unified multi-dimensional feature vector for use in subsequent models.

[0018] Step S20 in the method provided in this application embodiment includes: The time-domain features are obtained based on the time-domain waveform of the echo signal; The frequency domain features are extracted by performing a frequency domain transformation on the echo signal; Based on the complex data from each receiving channel, beamforming processing is performed to extract the spatial features, which include the beam pattern features after beamforming. Construct time-domain feature vectors, frequency-domain feature vectors, and spatial-domain feature vectors, normalize them, concatenate them in a predetermined order, perform feature fusion, and generate a multi-dimensional feature vector for input to the anomaly detection model.

[0019] In this embodiment, the time-domain waveform is the trajectory or shape of the amplitude of the echo signal received by the radar changing over time; the frequency-domain transformation is to convert the signal from the time dimension to the frequency dimension, and through the Fast Fourier Transform (FFT), the time-domain signal is decomposed into sinusoidal components of different frequencies to obtain the signal spectrum; the spectrum is a graph or data of the signal power distribution with frequency.

[0020] Specifically, firstly, calculations are performed on the original echo signal sequence without changing the data dimension. Through statistical and signal analysis methods, the overall energy level, peak characteristics, waveform stability, and fluctuation patterns of the sequence within a processing unit are described, and the calculated values ​​are the time-domain characteristics.

[0021] Secondly, the echo signal is transformed in the frequency domain to extract spectral features. Through statistical analysis, the overall energy level, peak characteristics, waveform stability and fluctuation patterns within a processing unit are obtained, including average power and spectral symmetry.

[0022] Next, beamforming processing is performed based on the complex data of each receiving channel to extract spatial features, including the radiation pattern features after beamforming. The complex data of each receiving channel is the original I / Q data collected by each independent receiving antenna element or subarray of the phased array radar, which retains the amplitude and phase information of the signal.

[0023] Beamforming processing involves adjusting the amplitude and phase of complex data from each channel to enhance the coherent superposition of signals from a specific direction, while suppressing signals from other directions, thus forming a high-gain beam in space. The radiation pattern is the graphical data representing the gain response function of waves from different spatial directions after beamforming processing, expressed in an angular azimuth coordinate system.

[0024] Specifically, the system receives the raw component data collected by each receiving antenna element of the phased array radar, adjusts and processes the complex data of each receiving antenna element to obtain the spatial characteristics in the spatial dimension, and represents the data in different directions to obtain the radiation pattern of the spatial characteristics.

[0025] Finally, based on the extracted time-domain features, frequency-domain features, and spatial-domain features, corresponding processing is performed to construct corresponding time-domain feature sub-vectors, frequency-domain feature sub-vectors, and spatial-domain feature sub-vectors. Then, normalization processing such as error elimination is performed on the time-domain feature sub-vectors, frequency-domain feature sub-vectors, and spatial-domain feature sub-vectors to obtain effective sub-vector data. Subsequently, the three processed sub-vectors are concatenated in a predetermined order, and after feature fusion, a multi-dimensional feature vector is generated for input into the anomaly detection model, including three types of feature vectors: time-domain features, frequency-domain features, and spatial-domain features.

[0026] The three sub-vectors of time domain features, frequency domain features, and spatial domain features are fused to obtain a multi-dimensional feature vector, which can be [-92,-70,0.03,2.5], [80,-22,0.95,0.85], [0.2,0.98,5].

[0027] Step S20 in the method provided in this application embodiment further includes: Calculate the average power, peak power, pulse width variation rate, and amplitude fluctuation standard deviation of the echo signal within a coherent processing interval, extract the time-domain features, and construct a time-domain feature sub-vector. Perform a Fast Fourier Transform on the echo signal to extract its main lobe width, side lobe level, spectral symmetry, and energy proportion within a specific frequency band. Extract the frequency domain features to form a frequency domain feature sub-vector. Based on the complex data from each receiving channel of the phased array radar, the spatial response under the current beam pointing is obtained through a digital beamforming algorithm. The beam pointing accuracy, beam shape distortion coefficient, and phase consistency error are extracted, and the spatial features are extracted to form a spatial feature sub-vector.

[0028] In this embodiment, the coherent processing interval is a basic time unit for radar signal processing. Within this interval, the initial phases of multiple pulses emitted by the radar are coherent, and the phases of the received echo signals have a deterministic relationship, which facilitates coherent accumulation and Doppler processing. The average power is the value obtained by statistically averaging the echo signal power of all pulses and all range units of interest within the coherent processing interval, reflecting the average level of received energy during this time period.

[0029] Peak power is the maximum power of all echo signals within the coherent processing interval, reflecting the intensity of possible strong scattering points or transient interference; pulse width variation rate describes the degree of variation of the echo signal envelope between pulses or between distance units within the coherent processing interval, usually calculated by calculating the statistical variance or a specific ratio of the width sequence; amplitude fluctuation standard deviation is a statistical measure of the deviation of the echo signal amplitude sequence from its average amplitude within the coherent processing interval.

[0030] First, the average power, peak power, pulse width variation rate, and amplitude fluctuation standard deviation of the echo signal within a coherent processing interval are calculated to extract time-domain features and form a time-domain feature sub-vector.

[0031] Specifically, within a coherent processing interval, the number of pulses and the I / Q data of all distance units within the coherent processing interval are read, and the instantaneous power of each unit is calculated. The number of instantaneous powers is the product of the number of pulses and the number of distance units.

[0032] Instantaneous power = I 2 +Q 2 Then, all instantaneous power values ​​are calculated, and the arithmetic mean of all power values ​​is obtained. Assume the calculated average is -92 dBm, representing the average echo intensity at that azimuth. The maximum value among all calculated power values ​​is then identified and taken as the peak power. This peak power may be due to ship targets or strong deceptive interference pulses. I / Q data represents the amplitude and phase information of a radio frequency signal in digital form through a pair of orthogonal baseband signals. Any sinusoidal signal is represented as a complex number, where I represents the in-phase component (the real part of the complex number), and Q represents the quadrature component (the imaginary part of the complex number).

[0033] For each coherent processing interval, the width of the signal envelope is analyzed. Then, the standard deviation of all signal envelope widths is calculated and divided by the mean to obtain the pulse width variation rate. The smaller the pulse width variation rate, the higher the stability of the pulse width in the range dimension. Subsequently, the envelopes of all analyzed signals are arranged to obtain the echo signal envelope sequence for all range cells. The standard deviation of the echo signal envelope sequence for all range cells is calculated to obtain the standard deviation of the echo signal amplitude fluctuation.

[0034] For this coherent processing interval, the average power, peak power, pulse width variation rate, and amplitude fluctuation standard deviation can be generated. The values ​​calculated for each distance unit are taken as a set of data and arranged in a specific order to obtain the time-domain feature sub-vector.

[0035] Secondly, a Fast Fourier Transform (FFT) is performed on the echo signal to extract its main lobe width, side lobe level, spectral symmetry, and energy proportion within a specific frequency band. Frequency domain features are extracted and frequency domain feature sub-vectors are constructed. The Fast Fourier Transform (FFT) is an algorithm for calculating the Discrete Fourier Transform, which converts a time-domain discrete signal into a frequency-domain discrete representation. The signal spectrum can be obtained through calculation. The main lobe width of the spectrum is the frequency width corresponding to the main peak in the power spectrum where the power drops to half of the peak value, reflecting the frequency resolution or potential distance resolution of the signal.

[0036] The sidelobe level of the spectrum is the level of the local peaks other than the main lobe in the power spectrum. It is usually expressed in decibels relative to the main lobe peak and reflects the degree of signal energy leakage and the ability to suppress interference. Spectral symmetry is a measure of whether the power spectrum of a signal is symmetrical around its center frequency. It can be used to calculate the energy ratio or correlation coefficient within equal bandwidths on both sides of the center frequency. The energy proportion within a specific frequency band is the proportion of the total spectral energy of the signal that falls within a reasonable bandwidth centered on the theoretical transmission frequency.

[0037] Specifically, after performing a fast Fourier transform on the echo signal, a complex spectrum can be obtained. The main lobe width, side lobe level, spectral symmetry, and energy proportion within a specific frequency band can be extracted from the complex spectrum.

[0038] The main lobe width is extracted; it is the power level when the peak value drops to half. The main lobe width is obtained by finding the frequency points on both sides of the main lobe peak where the amplitude first falls below half the peak value. The main lobe width is inversely proportional to the effective pulse width; a narrower main lobe indicates higher distance resolution. For example, at the peak value (e.g., at 0Hz), corresponding to the baseband center, points where the power drops by 3dB are found on both sides. Measuring the frequency difference between these two points yields a main lobe width of 80Hz.

[0039] Next, the sidelobe level is calculated. In the complex spectrum, the sidelobes are the lower energy peak regions on both sides of the main lobe. High sidelobes can mask weak nearby targets, causing false alarms or missed alarms. Therefore, it is necessary to calculate the sidelobe level to measure the signal's anti-interference capability and leakage risk. Outside the main lobe region, the sidelobe peak with the highest amplitude is found to obtain the sidelobe level, and its value should be as small as possible. Assuming that the highest sidelobe power is 22dB lower than the main lobe, the sidelobe level is recorded as -22dB.

[0040] In an ideal linear time-invariant system, the corresponding echo spectrum should be highly symmetrical. Asymmetry often indicates the presence of nonlinearity, I / Q imbalance, DC offset, or interference at specific frequencies. Therefore, it is necessary to calculate the spectral symmetry. Since the FFT spectrum is conjugate symmetrical about 0Hz, the positive and negative frequency components can be compared. The total energy within a certain frequency range on both sides of the center frequency is divided by the larger value, and the ratio is taken as the spectral symmetry. A ratio of 1 indicates perfect symmetry, while a ratio of 0 indicates complete asymmetry. The closer the ratio is to 1, the better.

[0041] Next, calculate the energy percentage within a specific frequency band. The energy percentage within a specific frequency band is the proportion of signal energy to the total energy of the entire frequency band within a certain frequency range of interest. First, set a target frequency band as a reference value. Then, calculate the total energy of the specific frequency band and the total energy of the target frequency band. Calculate the ratio of the total energy of the specific frequency band to the total energy of the target frequency band. For example, calculate the ratio to 0.85.

[0042] The main lobe width, side lobe level, spectral symmetry, and energy percentage within a specific frequency band calculated from a complex spectrum are arranged in a specific order to extract frequency domain features and obtain frequency domain feature subvectors.

[0043] For example, the numerical representation of the frequency domain feature vector is: [80, -22, 0.95, 0.85].

[0044] Finally, based on the complex data from each receiving channel of the phased array radar, the spatial response under the current beam pointing is obtained through a digital beamforming algorithm. Beam pointing accuracy, beam shape distortion coefficient, and phase consistency error are extracted to form spatial feature sub-vectors. Specifically, the digital beamforming (DBF) algorithm is an algorithm that performs spatial filtering and beam pointing control on the array's received signals in the digital domain; the spatial response is the gain response function of the array antenna to incident signals from different spatial directions; the beam pointing accuracy is the deviation between the spatial angle corresponding to the actual maximum gain point of the beam and the theoretically required pointing angle; the beam shape distortion coefficient is an indicator of the degree of difference between the actual measured array pattern shape and the ideal pattern shape; and the phase consistency error is the error between the inherent phase offset of each receiving channel's receiving link and the ideal phase value during actual operation.

[0045] Specifically, firstly, using the coherent signals received by the phased array's multi-channel receivers, a beam in a specific direction is synthesized through a digital beamforming algorithm. Then, the array's radiation pattern under the current direction is calculated, and the synthesized beam output signal may also be obtained. Next, beam pointing accuracy, beam shape distortion coefficient, and phase consistency error are extracted from the processing results to complete the extraction of radar spatial characteristics. Finally, these values ​​are arranged in order to form a spatial characteristic sub-vector.

[0046] The coherent signal received by a phased array multi-channel is obtained by weighted summation of the complex signals received by each array element, resulting in coherent enhancement in the desired direction and suppression in the interference / noise direction. For a specific beam, the corresponding direction is acquired to obtain the current radiation pattern of the phased array. Subsequently, beam pointing accuracy, beam shape distortion coefficient, and phase consistency error are calculated from the synthesized beam.

[0047] Acquire baseband I / Q data from all receiving channels at a given moment, and use the DBF algorithm to calculate the orthogonally pointing array pattern. In the calculated pattern, find the azimuth angle corresponding to the point of maximum gain, and use this angle as the beam pointing accuracy error.

[0048] Subsequently, the measured radiation pattern is compared with the ideal radiation pattern, and the normalized correlation coefficient in the main lobe region is calculated. This correlation coefficient is used as the beam shape distortion coefficient, with a value ranging from 0 to 1, where a larger value indicates stronger consistency. Next, the phase consistency error is extracted. Before applying beamforming weights, the inherent phase difference of each channel relative to the reference channel is estimated, and then compared with the ideal phase difference. The root mean square value of the phase deviation of all channels is calculated to obtain the phase consistency error. Finally, spatial feature extraction is completed, resulting in a spatial feature sub-vector: beam pointing accuracy, beam shape distortion coefficient, and phase consistency error. At this point, the feature sub-vectors in the time, frequency, and spatial domains are constructed, and normalization and fusion processing will follow.

[0049] For example, after completing the spatial feature extraction, a spatial feature sub-vector can be obtained as: [0.2, 0.98, 5].

[0050] In this embodiment, features are extracted and fused in parallel from three physical dimensions: time domain, frequency domain, and spatial domain. Then, anomaly indicators for each dimension are calculated to enhance the interpretability of the features. All anomalous features are associated with a specific performance of the radar or external influencing factors, thereby improving the discriminative ability of the features. This is beneficial for distinguishing hardware failures, environmental interference, and abnormal target characteristics. Subsequently, a fused feature vector is constructed through feature fusion to improve the dimension of anomaly analysis and avoid single-dimensional analysis.

[0051] S30: Construct an anomaly detection model, input the multidimensional feature vector into the anomaly detection model, and output the anomaly detection result; In this embodiment, the anomaly detection model is a mathematical function or computational graph constructed by an algorithm. Its core function is to learn the distribution pattern of normal data in the feature space and determine whether new input data deviates from the pattern.

[0052] Specifically, an anomaly detection model is constructed and trained based on historical echo data. Then, each multi-dimensional feature vector generated in real time is input into the anomaly detection model. Internally, the anomaly detection model calculates and evaluates the input vectors based on its learned knowledge, ultimately outputting anomaly detection results with confidence levels.

[0053] Step S30 in the method provided in this application embodiment includes: Historical echo data and historical anomaly type data within a preset time period are acquired, preprocessed, and a historical echo dataset is obtained. Anomaly types are labeled on the preprocessed historical echo data to construct a sample anomaly type set. Extract the multidimensional feature vector from the set of anomaly types in the labeled samples to construct a training sample set; The anomaly detection model is trained to obtain the anomaly detection model.

[0054] In this embodiment of the application, the preset time period is a past period of time or a complete mission cycle; the historical echo data is the multi-channel raw I / Q data acquired by the phased array radar system in the past actual operation, including raw observation records of various normal operating states and abnormal situations that have occurred.

[0055] Historical anomaly type data is a record file or database entry aligned with the time of historical echo data, including information such as the type of abnormal event of the radar system within the corresponding time period, the time of occurrence, and the duration; preprocessing is a series of cleaning, formatting, enhancement, and standardization operations performed on the data to improve data quality and unify data specifications; historical radar echo dataset is a collection of preprocessed data.

[0056] Specifically, firstly, historical echo data for a preset time period is extracted from the phased array radar management system, and historical anomaly type data is obtained from the corresponding maintenance logs. Then, the acquired historical echo data is preprocessed by data cleaning to remove errors, inconsistencies, and missing data, ensuring data quality. Finally, the processed historical echo data and historical anomaly type data are organized to obtain a historical radar echo dataset.

[0057] Secondly, based on the anomaly types within the historical anomaly types, the preprocessed historical echo data is annotated. Historical echo data with anomalies are labeled, while those without anomalies are discarded, thus constructing a set of sample anomaly types corresponding to the historical echo data. Anomaly labeling can be based on the anomaly type of the historical echo data, with different labels applied to different anomaly types. Through fine-grained labeling, each data point with anomalies in the historical echo data is automatically assigned an anomaly label to indicate the category to which the sample belongs. The historical echo data consists of time-domain feature vectors, frequency-domain feature vectors, and spatial-domain feature vectors, which are normalized and concatenated to obtain a multi-dimensional feature vector.

[0058] Next, multi-dimensional feature vectors are extracted from the set of anomaly types of labeled samples, and the set of anomaly types of labeled samples and historical echo data are used as the training sample set.

[0059] Finally, using the training sample set obtained above, the anomaly detection model is trained to obtain an anomaly detection model that can perform anomaly detection. After training, the anomaly detection model is used to detect the current echo data.

[0060] Feature vectors from the training sample set are input into the model in batches. The model performs forward propagation and outputs the predicted anomaly type or anomaly confidence. The prediction result is then compared with the true label, and the loss function value is calculated. Next, the gradient of the loss function with respect to the parameters of each layer of the model is calculated using the backpropagation algorithm. Finally, the optimizer updates the model parameters based on the gradient. This process is iterated multiple times on all training data until the model performance stabilizes and reaches a satisfactory level on an independent validation set. After training, an anomaly detection model with anomaly detection and classification capabilities is obtained.

[0061] For example, the parameters of a multi-branch deep learning network comprising three sub-networks in the time, frequency, and spatial domains are initialized. Then, the model is trained using a training sample set. A batch of data is taken from the training set and input into the model each time. The model performs forward propagation, outputting the probability distribution of each sample belonging to normal or various anomalies, and calculating the cross-entropy loss between the predicted probability distribution and the true label. Through backpropagation and the Adam optimizer, all weights and bias parameters in the network are adjusted to reduce the loss. This process is repeated iteratively, traversing all training data multiple times. After a certain number of training epochs, the model performance is evaluated to prevent overfitting. Training stops when the model's performance on the validation set no longer improves and reaches a predetermined standard: anomaly detection accuracy > 95% and classification accuracy > 90%. All parameters of the model at this point are saved, resulting in the trained anomaly detection model, which is then used for anomaly detection and recognition tasks.

[0062] In step S30 of the method provided in this application embodiment, constructing an anomaly detection model includes: An anomaly detection model is constructed using machine learning algorithms, namely a multi-branch deep learning network model, which includes branch networks corresponding to the time domain, frequency domain, and spatial domain, respectively. The three feature extraction branches, corresponding to the time-domain feature vector, the frequency-domain feature vector, and the spatial-domain feature vector, are each composed of a fully connected layer or a one-dimensional convolutional layer. The feature fusion layer, connected to the ends of the three branches, is used to concatenate or weightedly fuse the high-dimensional features extracted from each branch. The classification output layer, connected to the feature fusion layer, is used to output anomaly detection results, which include anomaly confidence levels.

[0063] In this embodiment, the machine learning algorithm is an algorithm that can automatically learn patterns from data and use these patterns to predict or make decisions on new data; the branch deep learning network model has multiple parallel and independent input processing paths, each branch is dedicated to processing a specific type or source of data, and finally the processing results of each branch are fused to make a comprehensive decision; the time domain / frequency domain / spatial domain branch network is a multi-branch model that is used to process time domain feature vectors, frequency domain feature vectors and spatial domain feature vectors respectively.

[0064] Specifically, by using machine learning algorithms, three independent branch networks are constructed simultaneously: a time-domain branch network, a frequency-domain branch network, and a spatial-domain branch network. Each branch is allowed to use the network layers and parameters best suited to its data characteristics, and learn the mapping relationship from low-level features to high-level abstract concepts in each domain. This can avoid information confusion or mutual interference between features from different domains, laying the foundation for subsequent high-quality feature fusion.

[0065] Secondly, there are three feature extraction branches corresponding to the temporal, frequency, and spatial feature vectors. The core of each branch consists of a fully connected layer or a one-dimensional convolutional layer. Before the core layer, each branch contains an input layer to receive input data. However, due to differences in the type and quantity of input data, each branch's input layer includes different nodes. The core layer of each branch performs feature extraction through a fully connected layer or a one-dimensional convolutional layer. For the temporal branch network, the corresponding input data consists of four independent parameters; for example, the input layer includes four nodes, and the core can consist of two fully connected layers. The frequency branch network has strong local correlation, and its core can consist of two one-dimensional convolutional layers, with the corresponding input data consisting of four parameters; for example, the input layer includes four nodes. The spatial branch network also has independent parameters as its input data. Its core is similar to the temporal branch network, consisting of two fully connected layers, with the corresponding input data consisting of three parameters; for example, the input layer includes three nodes.

[0066] Secondly, the feature fusion layer connected to the ends of the three branches is used to concatenate or weightedly fuse the high-dimensional features extracted from each branch. The end of the branch is the position of the final feature output by each feature extraction branch after completing its internal layer-by-layer calculations.

[0067] Specifically, high-dimensional feature vectors obtained from multiple parallel branches are concatenated to form a single ultra-long vector; alternatively, a weighted fusion method is used, assigning appropriate weights to the outputs of each branch and then combining them to generate a fused feature vector. This feature fusion layer integrates temporal energy, frequency structure, and spatial performance to detect anomalies in subsequent input data.

[0068] Finally, the classification output layer, connected to the feature fusion layer, outputs the anomaly detection result, which includes anomaly confidence. The classification output layer is the last layer in the neural network responsible for generating the final prediction; the anomaly confidence is a probability estimate of the anomaly detection model's judgment of an anomaly. It is typically a value between 0 and 1, with higher values ​​indicating greater confidence from the model.

[0069] After training is complete, the model can be updated using reinforcement learning algorithms. For example, anomaly detection models can be updated based on reinforcement learning.

[0070] For example, during model training, after comparing the predicted probability distribution with the true labels, the correct detection results are used as positive feedback and rewarded in the reinforcement learning process; false negatives and false positives are used as negative feedback and penalized. The model parameters are then updated via policy gradients, with an update performed after each batch of new data is processed or after each feedback is received. This process is repeated multiple times to maximize the long-term accumulated reward. After updating the model parameters via policy gradients, the updated anomaly detection model is finally obtained.

[0071] The multi-dimensional feature vectors are input into the anomaly detection network. Then, the time-domain branch network, frequency-domain branch network, and spatial-domain branch network process the corresponding data respectively. After that, nonlinear transformation is performed through fully connected layers or one-dimensional convolutional layers. Then, the feature fusion layer performs fusion analysis of anomaly detection. Finally, the classification output layer outputs the anomaly detection results, which may include anomaly confidence or the corresponding probability of anomaly type.

[0072] In this embodiment, an anomaly detection model, namely a multi-branch deep learning network model, is constructed using machine learning algorithms. Supervised training enhances the generalization ability and stability of the anomaly detection model, and reinforcement learning algorithms are used to update it. Subsequently, the time-domain, frequency-domain, and spatial-domain branches of the multi-branch deep learning network model are used to identify features in each dimension in parallel, achieving in-depth and personalized feature extraction for each dimension and avoiding information confusion during early feature mixing. A feature fusion layer is then used for high-level information association and integration to perform anomaly detection, improving the accuracy of anomaly detection. The system outputs the anomaly confidence score and anomaly type probability, and uses the confidence score to evaluate the reliability of the anomaly detection, improving the scientific rigor of the anomaly detection process and quantifying the evaluation results.

[0073] S40: When the detection result is abnormal, the abnormality type is identified; In this embodiment, the anomaly type identification is based on determining the existence of an anomaly, further analyzing the characteristic patterns of the anomaly, and classifying it into several predefined fault or interference categories; the anomaly confidence is the output of the anomaly detection model, representing the probability that the current data belongs to an anomaly category.

[0074] Specifically, when the output anomaly detection result exceeds a certain preset judgment threshold, anomaly type recognition is initiated, the feature vectors that have been judged as anomalies are analyzed, the probability distribution of their anomaly categories is output, and the category with the highest probability is taken as the main recognition result.

[0075] Step S40 in the method provided in this application embodiment includes: When the anomaly confidence level exceeds the preset first threshold, the current radar data or status is determined to be abnormal, and subsequent type identification is triggered; otherwise, it is determined to be normal. When an anomaly is identified, the anomaly type and the probability of each type are output. The type with the highest probability is used as the primary identification result. If the highest probability is lower than the preset second threshold, it is marked as an unknown type anomaly.

[0076] In this embodiment, the anomaly confidence level is the output of the anomaly detection model, representing the probability that the current input data belongs to an abnormal state; the preset first threshold is a judgment standard pre-set by the system or operator according to the task.

[0077] Specifically, firstly, the anomaly confidence level is usually a value between 0 and 1. If the anomaly confidence level exceeds the preset first threshold, the current radar data or status is determined to be abnormal; if the anomaly confidence level is lower than or equal to the first threshold, the current status is determined to be normal, and then the next step of input data processing is carried out.

[0078] If the first threshold is set too high, alarms will only be triggered for anomalies that the model is extremely confident in, increasing the risk of missed alarms but minimizing false alarms. If the first threshold is set too low, alarms will be triggered for any suspicious signs, resulting in more false alarms and fewer missed alarms. Therefore, the first threshold should not be too high or too low, and can be set between 0.6 and 0.8. At the same time, operations and maintenance personnel can make adaptive adjustments to the first threshold according to the actual task stage.

[0079] For example, if the preset first threshold is 0.75 and the output anomaly confidence level is 0.82, 0.82 > 0.75, the condition is met. The system determines that the current state is abnormal. Subsequently, the anomaly type identification process is triggered to determine the anomaly type and output the anomaly probability.

[0080] Secondly, when an anomaly is identified, the anomaly detection model will output the identified anomaly type and the anomaly probability corresponding to each anomaly type. The type with the highest probability is used as the main identification result. If the highest probability is lower than the preset second threshold, it is marked as an unknown type anomaly.

[0081] In this embodiment of the application, the main identification result is the most likely abnormal type from the probability distribution of abnormal types; the preset second threshold is to evaluate whether the confidence level of the type identification result is high enough.

[0082] Specifically, when an anomaly is detected, type identification is triggered. For the current feature vector, the probability of each category is calculated. The highest probability value among the probabilities of each category is taken as the primary identification result. Then, the highest probability value is compared with a preset second threshold. If the highest probability value is greater than or equal to the second threshold, the type identification result is considered reliable, and the anomaly type corresponding to the highest probability value is finally output as the identification result. If the highest probability value is lower than the second threshold, the type identification result output by the anomaly detection model is unreliable, the primary identification result is not adopted, and this anomaly identification is recorded as an unknown type anomaly. The first threshold should not be too high or too low, and can be set between 0.7 and 0.8. At the same time, the operation and maintenance personnel can make adaptive adjustments to the second threshold according to the actual task stage.

[0083] For example, assuming a second threshold of 0.7 and a model confidence level of 0.82, when an anomaly is detected, the output probability distribution is: hardware failure: 0.10, environmental interference: 0.85, target characteristic anomaly: 0.05. The highest probability is 0.85. 0.85 > 0.7. Therefore, the anomaly type identification result output by the anomaly detection is considered reliable, and the anomaly type is: environmental interference.

[0084] In step S40 of the method provided in this application embodiment, the exception types include: When an anomaly is detected, the anomaly type is identified, which is divided into three types: hardware failure, environmental interference, and target characteristic anomaly.

[0085] In this embodiment, hardware failure is an anomaly caused by the performance degradation, damage, or failure of the radar system's own physical components, and is not affected by instantaneous changes in the external environment; environmental interference is an anomaly caused by unwanted signals or phenomena in the external electromagnetic or physical environment of the radar system, and has time-varying, directional, and specific spectral / spatial characteristics; target characteristic anomaly is an abnormal data manifestation caused by the special electromagnetic scattering characteristics, motion characteristics, or behavioral patterns of the target itself.

[0086] In this embodiment, a dual threshold is used to determine the reliability of the anomaly type identification results. The first threshold is used to determine whether there are false anomalies. If there are false anomalies, an alarm is triggered to reduce the false alarm rate of anomaly detection. At the same time, the alarm threshold can be flexibly adjusted according to task requirements to improve the ability to identify false anomalies. The second threshold is used to evaluate the reliability of the anomaly confidence of the type identification. If it is less than the second threshold, the current anomaly can be determined to be unreliable and an unknown type of anomaly, thereby improving the reliability and trustworthiness of the system and providing reliable input for the subsequent generation of targeted instructions.

[0087] S50: Based on the anomaly type identification result, generate corresponding system adjustment instructions, upload the anomaly information, and then issue and execute the system adjustment instructions. In this embodiment, the adjustment command is a parameter modification or mode switching command that can be understood and executed by the radar control system; uploading means sending the information of the local radar node to a higher-level command center; and distributing means receiving instructions from a superior or generating instructions locally and then transmitting them to the radar's execution mechanism.

[0088] Specifically, based on the anomaly type identification results, targeted system adjustment instructions are automatically generated. Simultaneously, as an information node, detailed information about the anomaly is uploaded to the network center or command center to achieve battlefield situational awareness sharing and higher-level decision-making. After completing the information reporting, or immediately upon receiving confirmation / correction instructions from higher authorities, the generated adjustment instructions are issued to the radar's execution units for execution.

[0089] Step S50 in the method provided in this application embodiment includes: In anomaly identification, false alarms are judged, and the continuity and rationality of the flight path are verified by combining the multi-frame point trace information of the target, eliminating false anomaly alarms caused by clutter residue or intermittent interference. If a hardware failure is identified, a degraded operation or channel reconstruction command is generated, and a maintenance alarm is reported. If environmental interference is identified, commands for frequency switching, beam nulling, or transmit power adjustment are generated to avoid or suppress the interference. If the target is identified as having abnormal characteristics, a command is generated to extend the beam dwell time, adjust the transmitted waveform parameters, or focus on tracking, in order to identify and confirm the target.

[0090] In this embodiment, false alarm judgment is performed after an anomaly is initially identified to avoid ineffective occupation of system resources and the generation of misleading alarms; multi-frame point information is the observation data obtained by the radar in detecting the same spatial unit or target in multiple consecutive scanning cycles; continuity means that the target points should be smoothly connected in the time series; rationality means that the target's motion pattern should conform to basic kinematic laws and battlefield common sense.

[0091] Specifically, firstly, for the anomaly identification results of the current coherent processing interval, the target point information accumulated in multiple scan frames in the spatial direction is combined. Then, the points are correlated to form or update a potential track, and it is determined whether the points of the track can be smoothly connected and whether the motion state is reasonable. If the verification finds that the points are sparse, discontinuous, have absurd motion state, or cannot form a stable track at all, then the anomaly alarm is determined to be caused by instantaneous clutter residue, such as wave crests or random noise bursts or short intermittent interference, such as lightning. This is a false anomaly alarm and should be rejected, and will not proceed to the subsequent command generation process. If the verification finds that the points are smooth and the motion state is reasonable and stable, then it is determined to be a normal anomaly alarm, and the subsequent alarm process is carried out.

[0092] For example, suppose that at an azimuth of 90 degrees and a distance of 50 kilometers, a certain CPI anomaly detection model outputs a high confidence score, and the type identification initially points to anomalies in target characteristics, triggering a false alarm detection. By retrieving the point records of the vicinity of the range cell over the past 10 scan cycles, it is found that only the current frame and the previous frame have faint points in this location. Track verification reveals that the distance difference between the two points is extremely large, and both continuity and reasonableness checks fail. Therefore, it is determined that this target characteristic anomaly alarm is a false alarm.

[0093] If a hardware failure is identified, a degraded operation or channel reconfiguration command is generated, and a maintenance alarm is reported. After confirming the failure of some hardware functions, the radar system adjusts itself, generating a degraded operation command to maintain basic detection missions at a lower level. For example, it may shut down the faulty transmitter module and reduce the overall transmit power, or switch to a wider beam and lower gain mode after some receiver channels fail, or generate a channel reconfiguration command to perform self-repair, reconfigure the remaining healthy receiver channels, reconstruct a near-healthy array pattern, and compensate for performance loss, temporarily maintaining equipment operation and promptly reporting to the next higher-level command center, while also sending a maintenance alarm.

[0094] If environmental interference is identified, commands for frequency switching, beam null alignment, or transmit power adjustment are generated to avoid or suppress the interference. Upon encountering environmental interference, commands are generated to change the center frequency of the radar's transmitted signal, avoiding the frequency band where the interference signal is concentrated; alternatively, beam null alignment commands are generated on the array pattern to form extremely low-gain nulls with the direction of arrival of the interference signal, thereby significantly suppressing the reception of interference signals in that direction spatially; transmit power adjustment commands can also be generated to dynamically adjust the radar's transmit power. For example, when encountering suppressive interference, the transmit power may be appropriately reduced to decrease radar exposure; when penetration interference is required, power management may be employed, increasing the transmit power at critical moments. This is achieved by using frequency switching to avoid interference or by using beam null alignment to weaken interference at the receiver.

[0095] If an abnormal target characteristic is identified, commands to extend the beam dwell time, adjust the transmitted waveform parameters, or focus on tracking are generated for target identification and confirmation. When an abnormal target characteristic is detected, a command to extend the beam dwell time is generated, controlling the duration the radar beam remains at the suspected target's azimuth. This accumulates more echo pulses for the target, resulting in a higher signal-to-noise ratio and more refined feature data. Alternatively, commands to adjust the transmitted waveform parameters, such as modulation scheme, bandwidth, and pulse repetition frequency, are generated. For example, the simple pulse waveform used for searching is switched to a wideband linear frequency modulated waveform or phase-coded waveform used for identification to improve range resolution or anti-jamming capability. Or, a command to focus on tracking the target's trajectory is generated, allowing feature analysis or pattern matching to determine the target's type, attributes, or even identity.

[0096] For example, if a target has abnormal characteristics and a faint abnormal point is found at a certain azimuth elevation angle, which is suspected to be a special target, one of the following instructions is generated: the beam dwell time is extended from the usual 2 CPI to 10 CPI; the transmit waveform parameters are adjusted, switching from searching for waveforms with low pulse repetition frequency to identifying waveforms with medium pulse repetition frequency and medium bandwidth, and the instruction is tracked in a focused manner.

[0097] Step S50 in the method provided in this application embodiment further includes: When a single radar node identifies wide-area interference or complex anomalies, it uploads the anomaly information and preliminary analysis results to the command center via data link. Receive adjustment instructions from the command center, execute the instructions, and then report the results back to the command center.

[0098] In this embodiment, a single radar node is an independently operating radar station or radar platform; self-wide-area interference is an interference type that the radar itself identifies as having a wide coverage area and affecting multiple frequency points or airspace; composite anomaly is a situation where the radar simultaneously identifies multiple intertwined anomaly features, such as the coexistence of hardware failure and environmental interference features, or the simultaneous occurrence of multiple types of interference; data link is a communication network system for securely transmitting formatted digital information between different combat units; the command center is the information fusion and decision-making hub, responsible for receiving intelligence from various nodes, generating a global battlefield situation, and issuing instructions.

[0099] Specifically, firstly, when a single radar node identifies the current anomaly as either wide-area interference or a composite anomaly, the single radar node no longer relies on local resources for processing. Instead, it integrates the complete anomaly information of the event, including anomaly confidence, identified type, time of occurrence, azimuth, frequency, etc., as well as preliminary analysis results based on the local model, such as the rough characteristics of the interference and the suspected location of the fault, into a standard format report and uploads it to the command center via the data link network.

[0100] Secondly, after the command center analyzes and makes decisions, the radar node receives and parses the adjustment instructions issued by the command center, and makes adjustments accordingly. After executing the instructions, the results are fed back to the command center. For example, the instructions have been executed, the frequency has been switched to F3, the interference suppression effect is good, and the signal-to-noise ratio has improved by 10dB. If the feedback of the adjustment instructions is invalid, the command center can readjust the strategy and generate new instructions; if it is valid, no new instructions will be generated.

[0101] In this embodiment, the credibility of the command is improved by identifying false early warnings. Subsequently, targeted adjustment commands can be generated for different types of anomalies, thereby improving the radar's survivability and anomaly adjustment capabilities. Then, when a single radar node sends wide-area interference or composite anomalies, it reports, receives, and executes commands from the command center, and finally feeds back the execution results. This optimizes the radar's collaborative anomaly adjustment capabilities and improves the working efficiency of the radar equipment.

[0102] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, firstly, by acquiring multi-channel raw echo data in real time, the timeliness of anomaly detection and identification is ensured. From the multi-channel raw echo data, the amplitude and phase information of the received signals of each unit of the array antenna are obtained, providing data basis for subsequent multi-dimensional analysis in the time domain, frequency domain, and spatial domain, and laying a solid data foundation for anomaly detection.

[0103] Secondly, features are extracted and fused in parallel from three physical dimensions: time domain, frequency domain, and spatial domain. Then, anomaly indicators for each dimension are calculated to enhance the interpretability of the features. All anomalous features are associated with a specific performance of the radar or external influencing factors, thereby improving the discriminative ability of the features. This is beneficial for distinguishing hardware failures, environmental interference, and abnormal target characteristics. Subsequently, through feature fusion, a fused feature vector is constructed to increase the dimension of anomaly analysis and avoid single-dimensional analysis.

[0104] Secondly, an anomaly detection model, namely a multi-branch deep learning network model, is constructed using machine learning algorithms. Supervised training enhances the generalization ability and stability of the anomaly detection model, and reinforcement learning algorithms are used to update it. Subsequently, the time-domain, frequency-domain, and spatial-domain branches of the multi-branch deep learning network model are used in parallel to identify features across various dimensions, achieving multi-dimensional feature extraction and avoiding information confusion during early feature mixing. Then, a feature fusion layer is used for high-level information association and integration to perform anomaly detection, improving the accuracy of anomaly detection. The anomaly confidence score and anomaly type probability are output, and the confidence score is used to evaluate the reliability of the anomaly detection, improving the scientific rigor of the anomaly detection process and quantifying the evaluation results.

[0105] Meanwhile, the reliability of anomaly type identification results is judged by dual thresholds. The first threshold is used to determine whether there are false anomalies. If false anomalies are found, an alarm is triggered to reduce the false alarm rate of anomaly detection. The alarm threshold can be flexibly adjusted according to task requirements to improve the ability to identify false anomalies. The second threshold is used to evaluate the reliability of the anomaly confidence of type identification. If it is less than the second threshold, the current anomaly can be determined to be unreliable and an unknown type of anomaly, thereby improving the reliability and trustworthiness of the system and providing reliable input for the subsequent generation of targeted instructions.

[0106] Ultimately, the credibility of commands is improved by identifying false early warnings. Subsequently, targeted adjustment commands can be generated based on the type of anomaly, enhancing the radar's survivability and anomaly adjustment capabilities. Then, when a single radar node sends wide-area interference or complex anomalies, it reports, receives, and executes commands from the command center, and finally feeds back the execution results. This optimizes the radar's collaborative anomaly adjustment capabilities and improves the working efficiency of radar equipment.

[0107] Example 2, as Figure 2 As shown, based on the same inventive concept as the intelligent detection and identification method for radar anomaly data provided in Embodiment 1, this embodiment of the invention also provides an intelligent detection and identification system for radar anomaly data, the system comprising: Data acquisition module 11 is used to acquire multi-channel raw echo data from the phased array radar system in real time; Feature extraction module 12 is used to extract multi-dimensional feature vectors from raw echo data, wherein the extracted multi-dimensional feature vectors include time-domain features, frequency-domain features and spatial-domain features, and the spatial-domain features include the beam pattern features after beamforming; Anomaly detection module 13 is used to construct anomaly detection model, input the multidimensional feature vector into the anomaly detection model, and output anomaly detection results; The anomaly classification module 14 identifies the anomaly type when the detection result is abnormal. The instruction execution module 15 generates corresponding system adjustment instructions based on the anomaly type identification results, uploads the anomaly information, and then issues and executes the system adjustment instructions.

[0108] In one embodiment, the feature extraction module 12 is used for: The time-domain features are obtained based on the time-domain waveform of the echo signal; The frequency domain features are extracted by performing a frequency domain transformation on the echo signal; Based on the complex data from each receiving channel, beamforming processing is performed to extract the spatial features, which include the beam pattern features after beamforming. Construct time-domain feature vectors, frequency-domain feature vectors, and spatial-domain feature vectors, normalize them, concatenate them in a predetermined order, perform feature fusion, and generate a multi-dimensional feature vector for input to the anomaly detection model.

[0109] In one embodiment, the feature extraction module 12 is used for: Calculate the average power, peak power, pulse width variation rate, and amplitude fluctuation standard deviation of the echo signal within a coherent processing interval, extract the time-domain features, and construct a time-domain feature sub-vector. Perform a Fast Fourier Transform on the echo signal to extract its main lobe width, side lobe level, spectral symmetry, and energy proportion within a specific frequency band. Extract the frequency domain features to form a frequency domain feature sub-vector. Based on the complex data from each receiving channel of the phased array radar, the spatial response under the current beam pointing is obtained through a digital beamforming algorithm. The beam pointing accuracy, beam shape distortion coefficient, and phase consistency error are extracted, and the spatial features are extracted to form a spatial feature sub-vector.

[0110] In one embodiment, the anomaly detection module 13 is used for: Historical echo data and historical anomaly type data within a preset time period are acquired, preprocessed, and a historical echo dataset is obtained. Anomaly types are labeled on the preprocessed historical echo data to construct a sample anomaly type set. Extract the multidimensional feature vector from the set of anomaly types in the labeled samples to construct a training sample set; The anomaly detection model is trained to obtain the anomaly detection model.

[0111] The construction of the anomaly detection model includes: An anomaly detection model is constructed using machine learning algorithms, namely a multi-branch deep learning network model, which includes branch networks corresponding to the time domain, frequency domain, and spatial domain, respectively. The three feature extraction branches, corresponding to the time-domain feature vector, the frequency-domain feature vector, and the spatial-domain feature vector, are each composed of a fully connected layer or a one-dimensional convolutional layer. The feature fusion layer, connected to the ends of the three branches, is used to concatenate or weightedly fuse the high-dimensional features extracted from each branch. The classification output layer, connected to the feature fusion layer, is used to output anomaly detection results, which include anomaly confidence levels.

[0112] In one embodiment, the anomaly classification module 14 is used for: When the anomaly confidence level exceeds the preset first threshold, the current radar data or status is determined to be abnormal, and subsequent type identification is triggered; otherwise, it is determined to be normal. When an anomaly is identified, the anomaly type and the probability of each type are output. The type with the highest probability is used as the primary identification result. If the highest probability is lower than the preset second threshold, it is marked as an unknown type anomaly.

[0113] Among them, the exception types include: When an anomaly is detected, the anomaly type is identified, which is divided into three types: hardware failure, environmental interference, and target characteristic anomaly.

[0114] In one embodiment, the instruction execution module 15 is used to: In anomaly identification, false alarms are judged, and the continuity and rationality of the flight path are verified by combining the multi-frame point trace information of the target, eliminating false anomaly alarms caused by clutter residue or intermittent interference. If a hardware failure is identified, a degraded operation or channel reconstruction command is generated, and a maintenance alarm is reported. If environmental interference is identified, commands for frequency switching, beam nulling, or transmit power adjustment are generated to avoid or suppress the interference. If the target is identified as having abnormal characteristics, a command is generated to extend the beam dwell time, adjust the transmitted waveform parameters, or focus on tracking, in order to identify and confirm the target.

[0115] In one embodiment, the instruction execution module 15 is further configured to: When a single radar node identifies wide-area interference or complex anomalies, it uploads the anomaly information and preliminary analysis results to the command center via data link. Receive adjustment instructions from the command center, execute the instructions, and then report the results back to the command center.

[0116] Compared to existing technologies, this application firstly ensures the timeliness of anomaly detection and identification by acquiring multi-channel raw echo data in real time. From the multi-channel raw echo data, the amplitude and phase information of the received signals of each element of the array antenna are obtained, providing data basis for subsequent multi-dimensional analysis in the time domain, frequency domain, and spatial domain, and laying a solid data foundation for anomaly detection.

[0117] Secondly, features are extracted and fused in parallel from three physical dimensions: time domain, frequency domain, and spatial domain. Then, anomaly indicators for each dimension are calculated to enhance the interpretability of the features. All anomalous features are associated with a specific performance of the radar or external influencing factors, thereby improving the discriminative ability of the features. This is beneficial for distinguishing hardware failures, environmental interference, and abnormal target characteristics. Subsequently, through feature fusion, a fused feature vector is constructed to increase the dimension of anomaly analysis and avoid single-dimensional analysis.

[0118] Secondly, an anomaly detection model, namely a multi-branch deep learning network model, is constructed using machine learning algorithms. Supervised training enhances the generalization ability and stability of the anomaly detection model, and reinforcement learning algorithms are used to update it. Subsequently, the time-domain, frequency-domain, and spatial-domain branches of the multi-branch deep learning network model are used in parallel to identify features across various dimensions, achieving multi-dimensional feature extraction and avoiding information confusion during early feature mixing. Then, a feature fusion layer is used for high-level information association and integration to perform anomaly detection, improving the accuracy of anomaly detection. The anomaly confidence score and anomaly type probability are output, and the confidence score is used to evaluate the reliability of the anomaly detection, improving the scientific rigor of the anomaly detection process and quantifying the evaluation results.

[0119] Meanwhile, the reliability of anomaly type identification results is judged by dual thresholds. The first threshold is used to determine whether there are false anomalies. If false anomalies are found, an alarm is triggered to reduce the false alarm rate of anomaly detection. The alarm threshold can be flexibly adjusted according to task requirements to improve the ability to identify false anomalies. The second threshold is used to evaluate the reliability of the anomaly confidence of type identification. If it is less than the second threshold, the current anomaly can be determined to be unreliable and an unknown type of anomaly, thereby improving the reliability and trustworthiness of the system and providing reliable input for the subsequent generation of targeted instructions.

[0120] Ultimately, the credibility of commands is improved by identifying false early warnings. Subsequently, targeted adjustment commands can be generated based on the type of anomaly, enhancing the radar's survivability and anomaly adjustment capabilities. Then, when a single radar node sends wide-area interference or complex anomalies, it reports, receives, and executes commands from the command center, and finally feeds back the execution results. This optimizes the radar's collaborative anomaly adjustment capabilities and improves the working efficiency of radar equipment.

Claims

1. A method for intelligent detection and identification of radar anomaly data, characterized in that, Includes the following steps: Real-time acquisition of multi-channel raw echo data from phased array radar systems; Extracting multidimensional feature vectors from raw echo data, including time-domain features, frequency-domain features, and spatial-domain features, wherein the spatial-domain features include the beam pattern features after beamforming; Construct an anomaly detection model, input the multidimensional feature vector into the anomaly detection model, and output the anomaly detection result; When the detection result is abnormal, the type of abnormality is identified; Based on the anomaly type identification results, corresponding system adjustment instructions are generated, the anomaly information is uploaded, and then the system adjustment instructions are issued and executed.

2. The method according to claim 1, characterized in that, Multidimensional feature vectors are extracted from the raw echo data, including time-domain features, frequency-domain features, and spatial-domain features. The time-domain features are obtained based on the time-domain waveform of the echo signal; The frequency domain features are extracted by performing a frequency domain transformation on the echo signal; Based on the complex data from each receiving channel, beamforming processing is performed to extract the spatial features, which include the beam pattern features after beamforming. Construct time-domain feature vectors, frequency-domain feature vectors, and spatial-domain feature vectors, normalize them, concatenate them in a predetermined order, perform feature fusion, and generate a multi-dimensional feature vector for input to the anomaly detection model.

3. The intelligent detection and identification method for radar anomaly data according to claim 2, characterized in that, Constructing time-domain, frequency-domain, and spatial-domain feature sub-vectors includes: Calculate the average power, peak power, pulse width variation rate, and amplitude fluctuation standard deviation of the echo signal within a coherent processing interval, extract the time-domain features, and construct a time-domain feature sub-vector. Perform a Fast Fourier Transform on the echo signal to extract its main lobe width, side lobe level, spectral symmetry, and energy proportion within a specific frequency band. Extract the frequency domain features to form a frequency domain feature sub-vector. Based on the complex data from each receiving channel of the phased array radar, the spatial response under the current beam pointing is obtained through a digital beamforming algorithm. The beam pointing accuracy, beam shape distortion coefficient, and phase consistency error are extracted, and the spatial features are extracted to form a spatial feature sub-vector.

4. The method according to claim 1, characterized in that, Construct an anomaly detection model, input the multidimensional feature vector into the anomaly detection model, and output the anomaly detection result, including: Historical echo data and historical anomaly type data within a preset time period are acquired, preprocessed, and a historical echo dataset is obtained. Anomaly types are labeled on the preprocessed historical echo data to construct a sample anomaly type set. Extract the multidimensional feature vector from the set of anomaly types in the labeled samples to construct a training sample set; The anomaly detection model is trained to obtain the anomaly detection model.

5. The method according to claim 4, characterized in that, Constructing an anomaly detection model, including: An anomaly detection model is constructed using machine learning algorithms, namely a multi-branch deep learning network model, which includes branch networks corresponding to the time domain, frequency domain, and spatial domain, respectively. The three feature extraction branches, corresponding to the time-domain feature vector, the frequency-domain feature vector, and the spatial-domain feature vector, are each composed of a fully connected layer or a one-dimensional convolutional layer. The feature fusion layer, connected to the ends of the three branches, is used to concatenate or weightedly fuse the high-dimensional features extracted from each branch. The classification output layer, connected to the feature fusion layer, is used to output anomaly detection results, which include anomaly confidence levels.

6. The method according to claim 1, characterized in that, When the detection result is abnormal, the abnormality type is identified, including: When the anomaly confidence level exceeds the preset first threshold, the current radar data or status is determined to be abnormal, and subsequent type identification is triggered; otherwise, it is determined to be normal. When an anomaly is identified, the anomaly type and the probability of each type are output. The type with the highest probability is used as the primary identification result. If the highest probability is lower than the preset second threshold, it is marked as an unknown type anomaly.

7. The method according to claim 6, characterized in that, Exception types include: When an anomaly is detected, the anomaly type is identified, which is divided into three types: hardware failure, environmental interference, and target characteristic anomaly.

8. The method according to claim 1, characterized in that, Based on the anomaly type identification results, corresponding system adjustment instructions are generated, including: In anomaly identification, false alarms are judged, and the continuity and rationality of the flight path are verified by combining the multi-frame point trace information of the target, eliminating false anomaly alarms caused by clutter residue or intermittent interference. If a hardware failure is identified, a degraded operation or channel reconstruction command is generated, and a maintenance alarm is reported. If environmental interference is identified, commands for frequency switching, beam nulling, or transmit power adjustment are generated to avoid or suppress the interference. If the target is identified as having abnormal characteristics, a command is generated to extend the beam dwell time, adjust the transmitted waveform parameters, or focus on tracking, in order to identify and confirm the target.

9. The method according to claim 1, characterized in that, The process of uploading abnormal information, followed by issuing and executing system adjustment instructions, includes: When a single radar node identifies wide-area interference or complex anomalies, it uploads the anomaly information and preliminary analysis results to the command center via data link. Receive adjustment instructions from the command center, execute the instructions, and then report the results back to the command center.

10. A radar anomaly data intelligent detection and identification system, characterized in that, The system is used to implement the intelligent detection and identification method for radar anomaly data according to any one of claims 1-9, and the system includes: The data acquisition module is used to acquire multi-channel raw echo data from the phased array radar system in real time; The feature extraction module is used to extract multi-dimensional feature vectors from the raw echo data. The extracted multi-dimensional feature vectors include time-domain features, frequency-domain features, and spatial-domain features. The spatial-domain features include the beam pattern features after beamforming. Anomaly detection module is used to construct anomaly detection model, input the multidimensional feature vector into the anomaly detection model, and output anomaly detection results; The anomaly classification module identifies the anomaly type when the detection result is abnormal. The instruction execution module generates corresponding system adjustment instructions based on the anomaly type identification results, uploads the anomaly information, and then issues and executes the system adjustment instructions.

Images