Radar jamming classification method and system based on attention mechanism, electronic device and storage medium

Abstract

The application provides a radar interference classification method and system based on an attention mechanism, an electronic device and a storage medium, and relates to the technical field of attention mechanisms. The application obtains a pulse description word data stream, filters out pulse data segments with significant characteristics, uses a timestamp marking method to position the data segments in time, divides the data segments into multiple pulse interval segments by a preset pulse threshold, and each segment corresponds to one or more pulse intervals. The segments are subjected to multistage processing to obtain a multidimensional feature vector, which is input into a pre-trained interference classification neural network model. The model extracts local space-time features using a convolutional neural network, captures long-range dependencies between features and enhances key features using a multi-head attention mechanism, outputs a probability distribution of the type of spaceflight radar interference through a Softmax classification layer, and finally determines the classification result according to the maximum probability in the distribution, thereby improving the accuracy of spaceflight radar interference classification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of attention mechanism in the field of aerospace, and particularly relates to a radar interference classification method and system based on an attention mechanism, an electronic device and a storage medium. BACKGROUND

[0002] In the field of aerospace, space situational awareness and on-orbit safety scenarios of spacecraft, spacecraft-borne space radars (such as space-borne warning space radars) and ground deep space exploration space radars are core means for perceiving long-distance space targets (spacecrafts, space debris, non-cooperative targets), and the authenticity of the signals directly affects the accuracy of space target identification, orbit prediction, threat assessment and on-orbit decision-making. In the aerospace environment, a non-cooperative party may generate space-specific airborne interference signals through electronic countermeasures, such as simulating the orbit parameter pulse signal of a specific spacecraft, forging the scattering characteristic pulse sequence of space debris, injecting false orbit maneuver pulse data, etc., in order to hide real space targets, create false situations or mislead on-orbit decisions, and therefore there is an urgent need for a technology that can accurately identify space radar deep interference signals in real time to ensure the safety of spacecrafts in orbit and the reliability of space situational awareness, and to avoid risks such as orbit collision and target misjudgment caused by interference signals.

[0003] At present, the mainstream scheme for classifying space radar interference is a method based on signal feature extraction and deep learning classification. This scheme extracts time domain features of space radar signals, such as the correlation between pulse repetition period and orbit period, frequency domain features, such as the Doppler shift after deep space propagation correction, scattering features, such as the pulse amplitude corresponding to the space radar cross section, etc. After constructing a feature vector, the model is input into a convolutional neural network or a recurrent neural network, etc. The model is trained to learn the differences between real space target signals and space interference signals, and to realize classification and detection of interference types.

[0004] However, the existing scheme has obvious defects: space-specific airborne interference technology can simulate the static features of real space target space radar signals, making it difficult to distinguish between true and false based on extracted time domain and frequency domain features; at the same time, space interference signals may change dynamically with the orbit maneuver of space targets, and the model trained based on static features has weak generalization ability for dynamic orbit interference patterns, and is easily evaded by new space interference means, resulting in a decrease in classification accuracy and difficulty in meeting the high reliability requirements of space radar interference classification in complex space environments. SUMMARY

[0005] The present application aims to provide a radar interference classification method and system based on an attention mechanism, an electronic device and a storage medium, to solve the problem of low accuracy of space radar interference classification caused by the difficulty of adapting to the correlation between the orbit period and the pulse feature, the failure to consider the influence of deep space propagation delay, and the difficulty in dealing with orbit maneuver interference in the prior art.

[0006] To solve the above technical problems, in a first aspect, the application provides a radar interference classification method based on an attention mechanism, comprising:

[0007] Obtaining a pulse description word data stream collected by a spacecraft-borne radar or a ground deep space exploration spacecraft radar, and screening a pulse data segment with a significant space interference feature from the pulse description word data stream;

[0008] Timing positioning the pulse data segment using an ephemeris timestamp marking method, dividing the pulse data segment after timing positioning into multiple pulse interval segments by adapting a preset pulse threshold of a spatial target motion period, and each pulse interval segment corresponding to one or more pulse intervals;

[0009] Performing a multi-level processing of the spacecraft scene adaptation on the multiple pulse interval segments to obtain a multi-dimensional feature vector;

[0010] Inputting the multi-dimensional feature vector into a pre-trained spacecraft radar interference classification neural network model, the model using a convolutional neural network to extract local spatiotemporal features, and using a multi-head attention mechanism to capture long-range dependencies between features and enhance key features, and outputting a probability distribution of a spacecraft radar interference type corresponding to the multi-dimensional feature vector through a Softmax classification layer;

[0011] Determining a spacecraft radar interference classification result according to the maximum probability in the probability distribution.

[0012] Optionally, the inputting of the multi-dimensional feature vector into the pre-trained spacecraft radar interference classification neural network model, the model using a convolutional neural network to extract local spatiotemporal features, and using a multi-head attention mechanism to capture long-range dependencies between features and enhance key features, and outputting a probability distribution of a spacecraft radar interference type corresponding to the multi-dimensional feature vector through a Softmax classification layer, comprises:

[0013] Based on a convolution layer of a convolutional neural network of the pre-trained interference classification neural network model, performing local feature extraction on the multi-dimensional feature vector through a sliding window to obtain a local feature sequence;

[0014] After introducing nonlinearity through a ReLU activation function of the convolutional neural network, performing dimension adjustment on the local feature sequence through a preset size max-pooling layer to compress the feature dimension and retain key local features to obtain an adjusted local feature sequence;

[0015] Mapping the local feature sequence to a first feature matrix of a preset dimension through a fully connected layer of the convolutional neural network, so that the first feature matrix adapts to the input requirements of the multi-head attention mechanism;

[0016] generate a position encoding vector matching the length of the first feature matrix, and add the position encoding vector to the first feature matrix to obtain a second feature matrix;

[0017] divide the second feature matrix into a plurality of feature subsequences, each of which corresponds to an attention head;

[0018] calculate the correlation degree value between each attention head and the corresponding feature subsequence;

[0019] weight and fuse the feature subsequences based on the correlation degree value to obtain a plurality of attention features, and concatenate the plurality of attention features into a fusion feature sequence;

[0020] based on the Softmax classification layer in the interference classification neural network model, by numerical conversion of the matching degree of the fusion feature sequence and each type of spaceflight radar jamming, the probability distribution of the spaceflight radar jamming type corresponding to the multi-dimensional feature vector is obtained.

[0021] Optionally, the calculation of the correlation degree value between each attention head and the corresponding feature subsequence comprises:

[0022] convert the feature subsequence into a first feature component and a second feature component, the feature dimensions of the first feature component and the second feature component being the same;

[0023] calculate the correlation strength values of each feature component in the first feature component and each feature component in the second feature component, and based on each correlation strength value, form a numerical set;

[0024] adjust each correlation strength value in the numerical set so that the adjusted correlation strength value is within a preset interval, and determine the correlation values between features at each position in the feature subsequence according to the adjusted correlation strength value;

[0025] weight and integrate the correlation values based on the preset weight parameters corresponding to each attention head to obtain the correlation degree value between each attention head and the corresponding feature subsequence.

[0026] Optionally, the adjustment of each correlation strength value in the numerical set so that the adjusted correlation strength value is within a preset interval, and the determination of the correlation values between features at each position in the feature subsequence according to the adjusted correlation strength value, comprises:

[0027] set the upper limit value and the lower limit value of the preset interval, and judge each correlation strength value in the numerical set;

[0028] If the judgment result indicates that the correlation strength value is greater than the upper limit value, the upper limit value is taken as the correlation strength value, if the judgment result indicates that the correlation strength value is less than the lower limit value, the lower limit value is taken as the correlation strength value, and if the judgment result indicates that the correlation strength value is between the upper limit value and the lower limit value, the correlation strength value is kept unchanged to obtain an adjusted value, and the adjusted value is positively correlated with the correlation relationship between the features in the feature subsequence;

[0029] Based on all the adjusted values, the correlation values between the features in the feature subsequence are determined.

[0030] Optionally, the space scene adaptive multi-stage processing of the plurality of pulse interval segments to obtain a multi-dimensional feature vector comprises:

[0031] The feature statistics of each pulse interval segment are obtained, and the corresponding statistical features include at least one of the following: average interval, interval standard deviation, interval minimum value, interval maximum value, and interval histogram distribution;

[0032] The similarity coefficient of the statistical features between adjacent pulse interval segments and the variation coefficient of the pulse intervals in the same pulse interval segment are calculated.

[0033] The statistical features, the similarity coefficient, and the variation coefficient are associated to form a multi-dimensional feature vector.

[0034] Optionally, the calculation of the similarity coefficient of the statistical features between adjacent pulse interval segments and the variation coefficient of the pulse intervals in the same pulse interval segment comprises:

[0035] Two pulse interval segments arranged in time sequence and connected in position in the pulse interval segment are taken as adjacent segments;

[0036] For each type, the statistical features of the same type in the adjacent segments are calculated to obtain a corresponding numerical difference value;

[0037] The numerical difference values of all types of statistical features are integrated to obtain the similarity coefficient between the adjacent segments.

[0038] The average size and the dispersion value of all pulse interval values contained in the same pulse interval segment are calculated, and based on the average size and the dispersion value, the variation coefficient in the same pulse interval segment is calculated.

[0039] Optionally, the filtering of the pulse data segment with the significant space interference feature from the pulse description word data stream comprises:

[0040] extracting a repetition interval of the pulse interference from the pulse description word data stream;

[0041] identifying periodicity and randomness types of the repetition interval;

[0042] based on the periodicity, calculating a proportion of the randomness type in all types of the repetition interval, and screening out a pulse data segment with a significant periodicity feature or a significant randomness feature according to the proportion.

[0043] In a second aspect, the present application provides a radar interference classification system based on an attention mechanism, comprising:

[0044] an acquisition module configured to acquire a pulse description word data stream to be processed collected by a spacecraft-borne radar or a ground deep space exploration radar, and screen out a pulse data segment with a significant space interference feature from the pulse description word data stream;

[0045] a division module configured to perform time sequence positioning on the pulse data segment by using an ephemeris timestamp marking method, divide the time sequence positioned pulse data segment into a plurality of pulse interval segments by adapting a preset pulse threshold of a space target motion period, and make each pulse interval segment correspond to one or more pulse intervals;

[0046] a processing module configured to perform a multi-level processing of a space scene adaptation on the plurality of pulse interval segments, and obtain a multi-dimensional feature vector;

[0047] an output module configured to input the multi-dimensional feature vector into a pre-trained space radar interference classification neural network model, the model uses a convolutional neural network to extract local space-time features, uses a multi-head attention mechanism to capture long-range dependency relationships between features and enhance key features, and outputs a probability distribution of a space radar interference type corresponding to the multi-dimensional feature vector through a Softmax classification layer;

[0048] a determination module configured to determine a space radar interference classification result according to a maximum probability in the probability distribution.

[0049] In a third aspect, the present application provides an electronic device, comprising:

[0050] a memory configured to store a computer program;

[0051] a processor configured to implement steps of the radar interference classification method based on the attention mechanism according to the first aspect when the computer program is executed.

[0052] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is capable of implementing steps of the radar interference classification method based on the attention mechanism according to the first aspect when the computer program is executed by a processor.

[0053] In the present application, a radar interference classification method based on an attention mechanism is provided, which comprises: acquiring a pulse description word data stream to be processed, and screening out pulse data segments with significant features from the pulse description word data stream; using a time stamp marking method to perform timing positioning on the pulse data segments, dividing the timing-positioned pulse data segments into multiple pulse interval segments through a preset pulse threshold, each pulse interval segment corresponding to one or more pulse intervals; performing multi-level processing on the multiple pulse interval segments to obtain a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a pre-trained interference classification neural network model, the interference classification neural network model using a convolutional neural network to extract local space-time features, and using a multi-head attention mechanism to capture long-range dependencies between features and enhance key features, and outputting a probability distribution of a space radar interference type corresponding to the multi-dimensional feature vector through a Softmax classification layer; and determining a space radar interference classification result according to the maximum probability in the probability distribution.

[0054] The radar interference classification method based on the attention mechanism provided in the present application screens out pulse data segments with significant space interference features from a pulse description word data stream to be processed collected by a spacecraft-borne radar or a ground deep space exploration radar; uses an ephemeris time stamp marking method to perform timing positioning on the pulse data segments, divides the timing-positioned pulse data segments into multiple pulse interval segments through a preset pulse threshold adapted to the motion period of a space target, each pulse interval segment corresponding to one or more pulse intervals; performs multi-level processing adapted to a space scene on the multiple pulse interval segments to obtain a multi-dimensional feature vector; inputs the multi-dimensional feature vector into a pre-trained space radar interference classification neural network model, the model using a convolutional neural network to extract local space-time features, and using a multi-head attention mechanism to capture long-range dependencies between features and enhance key features, and outputting a probability distribution of a space radar interference type corresponding to the multi-dimensional feature vector through a Softmax classification layer; and determining a space radar interference classification result according to the maximum probability in the probability distribution.

[0055] Further, the local feature sequence is extracted by using the convolution layer, the dimension is adjusted by the activation function and the pooling layer, and then mapped to a feature matrix suitable for the multi-head attention mechanism through the full connection layer. After adding the position encoding, the feature sub-sequence is divided, the correlation degree value of each attention head and the corresponding feature sub-sequence is calculated, the fusion feature sequence is obtained by weighted fusion, and finally the probability distribution is output through the Softmax classification layer. When calculating the correlation degree value, the feature component needs to be converted, the correlation strength value is calculated and adjusted. Through the combination of convolution neural network and multi-head attention mechanism, the local space-time features can be effectively extracted, the long-range dependence between features can be captured and the key features can be enhanced, and through the fine correlation degree calculation, the rationality of feature weighting is improved, thereby improving the accuracy of spaceborne radar jamming classification and the adaptability to complex jamming scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 A flowchart of a radar jamming classification method based on an attention mechanism provided by an embodiment of the present application;

[0058] Figure 2 A specific implementation diagram of a radar jamming classification method based on an attention mechanism provided by an embodiment of the present application;

[0059] Figure 3 A structure diagram of a radar jamming classification system based on an attention mechanism provided by an embodiment of the present application. DETAILED DESCRIPTION

[0060] To solve the problems of incomplete feature association capture, fixed association degree calculation and insufficient dynamic adaptability in the prior art, an embodiment of the present application provides a radar interference classification method based on an attention mechanism, which adopts the following design concept: selecting a part with obvious features from collected pulse signal data, positioning in time sequence and then dividing into multiple pulse interval segments, and then obtaining feature vectors containing multiple aspects of information through multi-level processing; then inputting the feature vectors into a pre-trained neural network model, which combines convolution processing and multi-head attention processing: the convolution processing is used to extract local space-time features, and the multi-head attention processing is used to capture long-term relationships between features and highlight key information, and finally outputting the probabilities of various interferences through a classification layer, and determining the interference type according to the maximum probability. This method can more comprehensively capture the relationships between features through multi-head attention processing, and the flexible processing method can adapt to the characteristics of different interference types, thereby improving the adaptability to complex interference environments and effectively solving the problems existing in the prior art.

[0061] To make the person skilled in the art better understand the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0062] The core of the present application is to provide a radar interference classification method based on an attention mechanism, and a specific implementation process of the method is shown in the flowchart as Figure 1 The method comprises the following steps.

[0063] S11, acquiring a pulse description word data stream to be processed collected by a spacecraft-borne radar or a ground deep space exploration radar, and screening a pulse data segment with significant space interference features from the pulse description word data stream.

[0064] The pulse description word data stream is a data set containing various information of the pulse signal received by the space radar, and the information can include the time, amplitude, frequency, etc. of the pulse occurrence; the pulse data segment with significant features is a continuous data part selected from the data stream, which has obvious differences with other data in terms of time interval, amplitude change, etc.

[0065] In the embodiment of the present application, the data receiving device is used to obtain the pulse description word data stream to be processed, which covers all the pulse signal information received by the spaceborne radar, such as the time and amplitude data of 1000 continuous pulses in a certain period of time. Then, the significant feature judgment condition (such as the amplitude exceeding 500 units) is set, and the data stream is checked in sections according to the condition, and the continuous data part meeting the condition is screened out. For example, in the 1000 pulses, the amplitudes of the 301st to 500th pulses are all 520 to 600 (exceeding 500) and continuously appear, and this part becomes the significant feature pulse data section.

[0066] S12, the pulse data section is time-positioned by using the ephemeris timestamp marking method, the pulse data section after time-positioning is divided into multiple pulse interval segments by using a preset pulse threshold value adapted to the motion period of the space target, and each pulse interval segment corresponds to one or more pulse intervals.

[0067] In the embodiment of the present application, the time stamp marking method is used to add time information to each pulse to clearly show the order of appearance; the time positioning is determined according to the time stamp; the preset pulse threshold value is a preset division standard (based on the pulse interval time); the pulse interval is the time difference between two adjacent pulses; the pulse interval segment is a continuous data part divided according to the threshold value, containing one or more pulse intervals.

[0068] In the embodiment of the present application, the time stamp marking method is used to add time information to each pulse to clearly show the order of appearance; the time positioning is determined according to the time stamp; the preset pulse threshold value is a preset division standard (based on the pulse interval time); the pulse interval is the time difference between two adjacent pulses; the pulse interval segment is a continuous data part divided according to the threshold value, containing one or more pulse intervals.

[0069] S13, the multiple pulse interval segments are subjected to multi-level processing of space scene adaptation, and a multi-dimensional feature vector is obtained.

[0070] In the embodiment of the present application, the time stamp marking method is used to add time information to each pulse to clearly show the order of appearance; the time positioning is determined according to the time stamp; the preset pulse threshold value is a preset division standard (based on the pulse interval time); the pulse interval is the time difference between two adjacent pulses; the pulse interval segment is a continuous data part divided according to the threshold value, containing one or more pulse intervals.

[0071] In the embodiments of the present application, first-level processing is performed on each pulse interval segment obtained in S12 to count basic features (such as average, minimum, and maximum pulse interval). For example, a certain segment contains 5 intervals (2, 3, 5, 4, and 6 seconds), the average is (2+3+5+4+6) ÷ 5 = 4 seconds, the minimum is 2 seconds, and the maximum is 6 seconds. Second-level processing is then performed to calculate the dispersion degree as 1.414. Finally, these features are integrated and arranged in sequence into a four-dimensional vector [4, 2, 6, 1.414].

[0072] In S14, the multi-dimensional feature vector is input into a pre-trained space radar jamming classification neural network model. The model uses a convolutional neural network to extract local space-time features, and uses a multi-head attention mechanism to capture long-range dependencies between features and enhance key features. A Softmax classification layer is used to output the probability distribution of the space radar jamming type corresponding to the multi-dimensional feature vector.

[0073] In S14, the multi-dimensional feature vector is input into a pre-trained space radar jamming classification neural network model. The model uses a convolutional neural network to extract local space-time features, and uses a multi-head attention mechanism to capture long-range dependencies between features and enhance key features. A Softmax classification layer is used to output the probability distribution of the space radar jamming type corresponding to the multi-dimensional feature vector.

[0074] In the embodiments of the present application, the multi-dimensional feature vector obtained in S13 is first input into a pre-trained model. The convolutional neural network in the model extracts local space-time features (such as the variation rule of the last 3 features). Then, the multi-head attention mechanism is used to process these features. Each attention head focuses on different associations (such as one focuses on the average time and dispersion degree, and the other focuses on the minimum time and variation rate). After integration, the key features are enhanced. Finally, the Softmax classification layer is used to convert the probability of each type of jamming. For example, the probability distribution of noise jamming 0.7, deception jamming 0.2, and combined jamming 0.1 is obtained.

[0075] In S15, the maximum probability in the probability distribution is used to determine the space radar jamming classification result.

[0076] In S15, the maximum probability in the probability distribution is used to determine the space radar jamming classification result.

[0077] In the embodiments of the present application, the probability distribution output by S14 is first obtained. Then, the maximum probability 0.7 is found, and the corresponding jamming type, such as noise jamming, is determined as the final classification result.

[0078] Figure 2A specific implementation schematic diagram of a radar interference classification method based on an attention mechanism provided for an embodiment of the present application, and the implementation process is similar to the description of Figure 1 The present application provides the following specific examples based on the contents of Figure 1 and Figure 2 A space radar system obtains a pulse descriptor word data stream within 10 minutes through a signal receiving module, records the occurrence time and amplitude of all pulses in this period of time. The skilled person sets the condition that the amplitude is greater than 500 units as a significant feature, and selects the 100th to 300th pulse as a significant feature data segment. Then use the time stamp marking method to clearly indicate the occurrence time of each pulse, set the preset pulse threshold value to 8 seconds, calculate the interval between adjacent pulses: the interval between the 100th and 101st pulses is 9 seconds, the interval between the 150th and 151st pulses is 10 seconds, and the interval between the 180th and 181st pulses is 12 seconds, so it is divided into 4 segments. The first segment is processed in multiple stages: the first stage is to calculate the average interval of 5 seconds and the minimum of 3 seconds; the second stage calculates the dispersion degree and the change rate to form a five-dimensional vector [5, 3, 7, 1.414, 1.5], and the remaining segments are processed in the same way to obtain 3 vectors. Input the 4 vectors into the pre-trained model, the convolutional neural network extracts local features, the multi-head attention mechanism enhances key features, and the Softmax classification layer outputs the probability distribution. Finally, compare the probabilities, the maximum probability 0.7 corresponds to noise interference, and determine the classification result.

[0079] By performing S11-S15, the embodiment of the present application reduces irrelevant data interference by selecting significant feature data segments from the pulse descriptor word data stream, and focuses on effective analysis objects; through time sequence positioning and dividing pulse interval segments, the pulse data is structured, laying a foundation for subsequent processing; the multi-dimensional feature vector obtained by multi-stage processing comprehensively captures the key characteristics of the pulse segment; combined with the model processing of convolutional neural network and multi-head attention mechanism, the key features are effectively extracted and the feature correlation is captured, and the output probability distribution more accurately reflects the interference possibility; finally, the classification result is determined according to the maximum probability, making the conclusion clear and intuitive. The whole process works together to improve the accuracy of space radar interference classification and the adaptability to complex environments, meeting the demand for accurately judging the type of interference in actual application.

[0080] In one possible embodiment, S14, input the multi-dimensional feature vector into the pre-trained space radar interference classification neural network model, the model uses convolutional neural network to extract local space-time features, and uses multi-head attention mechanism to capture long-range dependencies between features and enhance key features, and outputs the probability distribution of the space radar interference type corresponding to the multi-dimensional feature vector through the Softmax classification layer, including:

[0081] Step 141, based on the convolution layer of the pre-trained interference classification neural network model, local feature extraction is performed on the multi-dimensional feature vector through a sliding window to obtain a local feature sequence.

[0082] In the formula, the convolution layer is a part of the interference classification neural network model for extracting local features, the sliding window is a fixed-size data analysis range that can move over different regions of the data, the multi-dimensional feature vector contains multiple feature information, and the local feature sequence is a sequence composed of local features extracted from the multi-dimensional feature vector in sequence through the sliding window.

[0083] In the embodiment of the present application, based on the convolution layer of the pre-trained model, the size of the sliding window is set, and the window moves on the multi-dimensional feature vector at a certain step size. The integrated information (such as the average value) of the features in the window is calculated after each movement, and the local feature sequence is obtained in the order of movement. For example, the multi-dimensional feature vector [5, 3, 7, 1.414, 1.5], the window size is 3, and the step size is 1. The average values of the first three features are calculated in turn (5+3+7) ÷3=5, the middle three features are 3.805, and the last three features are 3.305, and the sequence [5, 3.805, 3.305] is obtained.

[0084] Step 142, after introducing nonlinearity through the ReLU activation function of the convolutional neural network, the feature dimension is compressed through a preset size of the max pooling layer to adjust the dimension of the local feature sequence to retain key local features, and an adjusted local feature sequence is obtained.

[0085] In the formula, the ReLU activation function is a processing method for introducing nonlinearity, the feature value is kept as the original value when it is greater than 0, otherwise it is 0; the max pooling layer of the preset size is a fixed-size screening range for selecting the maximum value from the local features; the compressed feature dimension is to reduce the sequence length, the key local feature is a feature more important for classification, and the adjusted local feature sequence is a sequence after compression and retention of key features.

[0086] In the embodiment of the present application, the local feature sequence of step 141 is input into the ReLU activation function to retain positive values; then the max pooling layer of the preset size is used to select the maximum value from the continuous features to form the adjusted sequence in order. For example, the sequence [5, 3.805, 3.305] is unchanged after ReLU processing, and the max pooling layer with a size of 2 is used to select the maximum values of the first two features 5 and the remaining feature 3.305 to obtain [5, 3.305].

[0087] Step 143, the local feature sequence is mapped to a first feature matrix of a preset dimension through the fully connected layer of the convolutional neural network, so that the first feature matrix adapts to the input requirements of the multi-head attention mechanism.

[0088] Wherein, the full connection layer is the part of integrating features and converting dimensions in the neural network, the local feature sequence is the sequence after dimension adjustment, the preset dimension is the matrix dimension required by the multi-head attention mechanism, the first feature matrix is the converted matrix, and the adaptation input requirement refers to that the dimension meets the processing standard of the multi-head attention mechanism.

[0089] In the embodiment of the application, the local feature sequence adjusted in step 142 is input into the full connection layer, and the sequence is converted into a first feature matrix of a preset dimension through internal calculation rules. For example, the sequence [5, 3.305] is converted into a 2-row 2-column matrix [[5, 2.1], [3.305, 1.8]], so as to adapt to the input of the multi-head attention mechanism.

[0090] Step 144, a position encoding vector matching the length of the first feature matrix is generated, and the position encoding vector is added to the first feature matrix to obtain a second feature matrix.

[0091] Wherein, the position encoding vector is a vector marking feature position information, and the length is consistent with the first feature matrix; the first feature matrix is a matrix adapted to the multi-head attention mechanism; and the second feature matrix is the result of addition of the two, used to retain position information.

[0092] In the embodiment of the application, a position encoding vector matching the length of the first feature matrix is generated, and the corresponding elements are added to obtain the second feature matrix. For example, the first feature matrix is [[5, 2.1], [3.305, 1.8]], the position encoding vector is [[0.1, 0.2], [0.3, 0.4]], and after addition, [[5.1, 2.3], [3.605, 2.2]] is obtained.

[0093] Step 145, the second feature matrix is divided into a plurality of feature subsequences, and each feature subsequence corresponds to an attention head.

[0094] Wherein, the second feature matrix is a matrix containing position information, the feature subsequence is a segmented subpart, and the attention head is a processing branch of the multi-head attention mechanism, and each attention head corresponds to a feature subsequence.

[0095] In the embodiment of the application, according to the number of attention heads, the second feature matrix is divided into a corresponding number of feature subsequences, and each subsequence is assigned to an attention head. For example, 2 attention heads divide the matrix [[5.1, 2.3], [3.605, 2.2]] into [[5.1, 2.3]] and [[3.605, 2.2]] by rows, which correspond to two attention heads respectively.

[0096] Step 146, the correlation degree value between each attention head and the corresponding feature subsequence is calculated.

[0097] Wherein, the attention head is a processing branch of a multi-head attention mechanism, the feature subsequence is a submatrix assigned to the attention head, and the correlation degree value is a numerical value measuring the close degree of correlation between the two.

[0098] In the embodiment of the present application, the feature subsequence is converted into two feature components of the same dimension, the correlation strength value of the corresponding elements is calculated (such as multiplication and summation), and after adjustment to the preset interval, the correlation degree value is obtained by combining and weighting the preset weight parameter of the attention head. For example, the subsequence [[5.1, 2.3]] is converted into two same components, the correlation strength is 5.1*5.1+2.3*2.3=31.3, which remains unchanged in the preset interval of 0 to 50, and is multiplied by the weight 0.8 to obtain 25.04.

[0099] Step 147, based on the correlation degree value, the feature subsequence is weighted and fused to obtain a plurality of attention features, and the plurality of attention features are spliced into a fusion feature sequence.

[0100] Wherein, the correlation degree value is a numerical value measuring the close degree of correlation, the weighted fusion is a process of adjusting the importance of the feature subsequence and merging according to the value, the attention feature is the feature processed by each attention head, and the fusion feature sequence is a sequence spliced by a plurality of attention features.

[0101] In the embodiment of the present application, the feature subsequence is weighted according to the correlation degree value of step 146 to obtain the attention feature of each attention head, and then spliced into a fusion feature sequence. For example, the first subsequence [[5.1, 2.3]] is weighted by the correlation degree value 25.04 to obtain [[127.7, 57.6]], the second subsequence [[3.605, 2.2]] is weighted by 18.5 to obtain [[66.7, 40.7]], and after splicing, it is [[127.7, 57.6], [66.7, 40.7]].

[0102] Step 148, based on the Softmax classification layer in the interference classification neural network model, the probability distribution of the spaceflight radar interference type corresponding to the multi-dimensional feature vector is obtained by numerically converting the matching degree of the fusion feature sequence and each type of spaceflight radar jamming.

[0103] Wherein, the Softmax classification layer is a part of the output probability distribution, the fusion feature sequence is the integrated sequence, the matching degree is the similarity degree of the sequence and each type of interference, the numerical conversion is the process of converting the matching degree to the probability, and the probability distribution is the possibility set of each type of interference.

[0104] In the embodiment of the present application, the fusion feature sequence of step 147 is input into the Softmax classification layer, the matching degrees with each type of interference are calculated, and then the matching degrees are converted into probabilities (each matching degree is divided by the sum) to form a probability distribution. For example, the matching degree with noise interference is 80, the matching degree with deception interference is 30, the sum is 110, and the converted probabilities are 0.73 and 0.27.

[0105] The present application provides the following specific examples: a space radar system processes a multi-dimensional feature vector [5, 3, 7, 1.414, 1.5]: step 141 uses a convolution layer with a window size of 3 and a step size of 1 to calculate the sum in the window to obtain a local feature sequence [15, 11.414, 9.914]; step 142 processes through ReLU, and then uses a maximum pooling layer with a size of 2 to select the maximum value to obtain an adjusted sequence [15, 9.914]; step 143 converts through a fully connected layer to obtain a first feature matrix of 2 rows and 2 columns [[15, 6.2], [9.914, 4.5]]; step 144 adds a position encoding vector [[0.5, 0.3], [0.2, 0.1]] to obtain a second feature matrix [[15.5, 6.5], [10.114, 4.6]]; step 145 is divided into [15.5, 6.5] and [10.114, 4.6] according to 2 attention heads; step 146 calculates the correlation strength (15.5*15.5+6.5*6.5=282.5 and 10.114*10.114+4.6*4.6≈121.8), respectively multiplied by the weights 0.7 and 0.85 to obtain the correlation degree values 197.75 and 103.5; step 147 obtains the attention features [3065.1, 1285.4] and [1065.0, 484.4] after weighting, and splices them into a fusion feature sequence; step 148 calculates through the Softmax classification layer, the matching degree with noise interference is 140, the matching degree with deception interference is 60, the sum is 200, and the converted probabilities are 0.7 and 0.3 to obtain a probability distribution.

[0106] By performing steps 141-148, the embodiment of the present application extracts local features through a convolution layer, optimizes feature dimensions through an activation function and a pooling layer, and adapts the input requirements of the multi-head attention mechanism through a fully connected layer; adding position encoding preserves the sequence information of the features, the multi-head attention mechanism integrates multi-dimensional feature correlations by dividing sub-sequences, calculating correlation degree values, and weighting fusion; finally, the probability distribution is output through the Softmax classification layer. This process not only captures local feature details, but also integrates long-range feature correlations, preserves key information, and quantifies the possibility of interference types, thereby improving the accuracy of space radar interference classification and the ability to process complex features.

[0107] In one possible embodiment, step 146, calculating the correlation degree value between each attention head and the corresponding feature sub-sequence, includes:

[0108] a1, convert the feature sub-sequence into a first feature component and a second feature component, the first feature component and the second feature component have the same feature dimension.

[0109] The feature sub-sequence is a sub-part segmented from the second feature matrix and used to correspond to a specific attention head. The first feature component and the second feature component are two feature parts obtained by converting the feature sub-sequence, and the two feature parts have the same number and arrangement of features, i.e., the same feature dimension.

[0110] In the embodiment of the present application, the feature sub-sequence obtained in step 145 is converted into two feature components by copying or mapping, one of which is the first feature component and the other is the second feature component, to ensure that the feature dimensions of the two components are completely the same. For example, the feature sub-sequence is [15.5, 6.5], and after copying, the first feature component is [15.5, 6.5] and the second feature component is [15.5, 6.5], both of which are 1 row and 2 columns of feature structure.

[0111] a2, calculate the correlation strength value of each feature component in the first feature component and each feature component in the second feature component, and based on the correlation strength value, form a numerical set.

[0112] The first feature component and the second feature component are two parts converted from the feature sub-sequence. The correlation strength value is a value for measuring the close degree of the correlation between the features at the corresponding positions in the two components. The numerical set is the whole formed by all the correlation strength values.

[0113] In the embodiment of the present application, the features at the corresponding positions in the first feature component and the second feature component are taken out, the correlation strength value is calculated by multiplying and summing, and then all the calculation results are arranged to form the numerical set. For example, the features 15.5 at the corresponding positions in the first feature component [15.5, 6.5] and the second feature component [15.5, 6.5] are multiplied to obtain 240.25, and the features 6.5 at the corresponding positions are multiplied to obtain 42.25. The sum of the two is 282.5, which is the correlation strength value. Thus, the numerical set is formed.

[0114] a3, adjust each correlation strength value in the numerical set so that the adjusted correlation strength value is within a preset interval, and determine the correlation value between the features at each position in the feature sub-sequence according to the adjusted correlation strength value.

[0115] The value set is an entirety of the correlation strength values, the preset interval is a preset value range for regulating the size of the correlation strength values, the adjusted correlation strength value is a value in the preset interval after processing, and the correlation value is a value determined according to the adjusted correlation strength value and representing the correlation degree between the features at each position in the feature subsequence.

[0116] In the embodiment of the application, the preset interval is determined, and whether each correlation strength value in the value set is in the interval is checked one by one. If the value is less than the lower limit of the interval, the value is adjusted to the lower limit value, if the value is greater than the upper limit of the interval, the value is adjusted to the upper limit value, and if the value is in the interval, the value remains unchanged. The adjusted correlation strength value is the correlation value between the features at each position in the feature subsequence. For example, the preset interval is 0 to 500, and the value 282.5 in the value set is in the interval, so the value remains 282.5 after adjustment. The value is the correlation value.

[0117] a4, based on the preset weight parameter corresponding to each attention head, weighting and integrating the correlation value to obtain the correlation degree value between each attention head and the corresponding feature subsequence.

[0118] The attention head is a processing branch in the multi-head attention mechanism, each attention head has a corresponding preset weight parameter, the parameter is used to measure the importance of the attention head in the overall processing, the correlation value is a value representing the correlation degree between the features in the feature subsequence, and the correlation degree value is a final value obtained by weighting and integrating, which measures the correlation degree between the attention head and the corresponding feature subsequence.

[0119] In the embodiment of the application, the preset weight parameter corresponding to each attention head is obtained, the correlation value obtained in step a3 is multiplied by the weight parameter, and the result is the correlation degree value between the attention head and the corresponding feature subsequence. For example, the correlation value is 282.5, the preset weight parameter is 0.7, and 282.5 is multiplied by 0.7 during calculation to obtain 197.75. The result is the correlation degree value.

[0120] The application provides the following specific examples: the feature subsequence corresponding to the first attention head in the aerospace radar system is [15.5, 6.5], the subsequence is copied through the a1 step first, to obtain the first feature component [15.5, 6.5] and the second feature component [15.5, 6.5]; when the a2 step calculates the correlation strength value, the features at the positions corresponding to the two components are multiplied respectively, 15.5*15.5=240.25, 6.5*6.5=42.25, then the two products are added to obtain 240.25+42.25=282.5, to form a value set; the a3 step sets the preset interval as 0 to 500, and 282.5 is in the interval, so it is directly determined as the correlation value; the preset weight parameter of the attention head in the a4 step is 0.7, and the correlation value 282.5 is multiplied by 0.7, that is, 282.5*0.7=197.75, and finally the correlation degree value 197.75 is obtained.

[0121] By performing a1-a4, the embodiment of the application provides a consistent basis for correlation calculation by converting the feature subsequence into two components of the same dimension; the quantification of feature correlation is realized by calculating the correlation strength value and forming a set; the adjustment of the value to the preset interval avoids the interference of extreme values on subsequent processing; the correlation degree value obtained in combination with the preset weight parameter enables the correlation measure to reflect the importance of the attention head. This series of operations systematically completes the quantification of the correlation degree of the attention head and the feature subsequence, provides a standard, reasonable and accurate basis for subsequent feature weighted fusion, and helps to improve the effectiveness of the overall processing process.

[0122] In a possible embodiment, a3, each correlation strength value in the value set is adjusted so that the adjusted correlation strength value is within the preset interval, and the correlation value between the features at each position in the feature subsequence is determined according to the adjusted correlation strength value, including:

[0123] b1, set the upper limit value and the lower limit value of the preset interval, and judge each correlation strength value in the value set.

[0124] Wherein, the upper limit value and the lower limit value of the preset interval are two values preset to delimit the reasonable range of the correlation strength value, the upper limit value is the maximum value of the interval, and the lower limit value is the minimum value of the interval; the value set is a whole composed of multiple correlation strength values, which are used to reflect the correlation closeness between the features in the feature subsequence; the correlation strength value is a specific value for measuring the correlation degree between the features.

[0125] In this embodiment of the application, an upper limit and a lower limit of a preset range are set according to actual needs. For example, the upper limit is set to 500 and the lower limit is set to 0. Then, each correlation strength value is taken out from the value set and compared with the upper limit and the lower limit one by one to determine whether each value is greater than the upper limit, less than the lower limit, or in between. For example, if the value set is [282.5, 550, -30], it is determined that 282.5 is between 0 and 500, 550 is greater than 500, and -30 is less than 0.

[0126] b2. If the judgment result indicates that the correlation strength value is greater than the upper limit value, then the upper limit value is used as the correlation strength value. If the judgment result indicates that the correlation strength value is less than the lower limit value, then the lower limit value is used as the correlation strength value. If the judgment result indicates that the correlation strength value is between the upper limit value and the lower limit value, then the correlation strength value is kept unchanged to obtain the adjusted value. The adjusted value is positively correlated with the correlation relationship between each feature in the feature subsequence.

[0127] The judgment result is the conclusion obtained by comparing the correlation strength value with the upper and lower limits of the preset interval; the adjusted value is the value of the correlation strength value after processing and falling within the preset interval; positive correlation means that the larger the adjusted value, the closer the correlation between the features in the feature subsequence, and the smaller the value, the looser the correlation.

[0128] In this embodiment of the application, the correlation strength value is adjusted according to the judgment result of step b1. If the value is greater than the upper limit, the upper limit value is used to replace the value; if the value is less than the lower limit, the lower limit value is used to replace the value; if the value is between the upper and lower limits, the original value is kept unchanged, and the adjusted value is obtained. The adjusted value is positively correlated with the feature. For example, in the value set [282.5, 550, -30], 550 is adjusted to 500, -30 is adjusted to 0, and 282.5 remains unchanged, resulting in [282.5, 500, 0]. The larger the value, the stronger the feature correlation.

[0129] b3. Based on all adjusted values, determine the correlation values ​​between features in the feature subsequence.

[0130] The adjusted values ​​are those that have been processed within a preset range and are within a reasonable range. They are used to standardize the degree of correlation between features. The correlation values ​​between features in the feature subsequence are determined based on all the adjusted values ​​and are used to finally quantify the degree of correlation between features.

[0131] In the embodiment of the present application, all the adjusted values obtained in the b2 step are collected, which are in the preset interval and reflect the correlation degree between the features, and the adjusted values are directly determined as the correlation values between the features in the feature subsequence. For example, the adjusted values are [282.5, 500, 0], and the corresponding correlation values are [282.5, 500, 0].

[0132] The present application provides the following specific examples: the space radar system processes the correlation strength value set [282.5, 550, -30] of the feature subsequence: the b1 step sets the upper limit value 500 and the lower limit value 0 of the preset interval, and each value is judged one by one. 282.5 is between 0 and 500, 550 is greater than 500, and -30 is less than 0; the b2 step adjusts the values according to the judgment result. 550 is replaced by 500 because it is greater than the upper limit value, -30 is replaced by 0 because it is less than the lower limit value, and 282.5 remains unchanged because it is in the interval. The adjusted values [282.5, 500, 0] are obtained, and the greater the values, the closer the correlation between the corresponding features.

[0133] By performing b1-b3, the embodiment of the present application provides a clear standard for the adjustment of the correlation strength value by setting the upper and lower limit values of the preset interval; the targeted adjustment based on the relationship between the value and the interval can avoid the interference of extreme values on the correlation measurement while maintaining the positive correlation between the value and the feature correlation; the adjusted value is determined as the correlation value, which can standardize and accurately quantify the correlation degree between the features, provide a reliable data basis for subsequent processing, and improve the rationality and effectiveness of the correlation analysis.

[0134] In a possible embodiment, S13, the multi-level processing of the space scene adaptation of the plurality of pulse interval segments, obtains a multi-dimensional feature vector:

[0135] Step 131, performing feature statistics on each pulse interval segment to obtain corresponding statistical features including at least one of the following: average interval, interval standard deviation, interval minimum value, interval maximum value, and interval histogram distribution.

[0136] Wherein, the pulse interval segment is a continuous pulse interval data part divided by time, the feature statistics is a calculation and analysis of the pulse interval in the segment to obtain data reflecting its characteristics, and the statistical features include the average interval, the interval standard deviation, the interval minimum value, the interval maximum value, and the interval histogram distribution. These features collectively describe the basic characteristics of the pulse interval segment.

[0137] In the embodiments of the present application, each pulse interval segment is selected, and all pulse intervals in the segment are characterized. The average interval is calculated by adding all pulse interval values and dividing by the number of intervals. For example, a segment has 5 pulse intervals of 2 seconds, 3 seconds, 5 seconds, 4 seconds, and 6 seconds, and the average interval is (2+3+5+4+6) ÷ 5 = 4 seconds. The interval standard deviation is calculated by first calculating the difference between each interval and the average interval, squaring the values, averaging the squared values, and then taking the square root. For the example, the difference is 2, the squared value is 4, the average is 2, and the square root is approximately 1.414. The minimum interval of 2 seconds and the maximum interval of 6 seconds are recorded, and the histogram distribution is counted, for example, 2-3 seconds appear twice, and 4-6 seconds appear three times.

[0138] Step 132, calculating the similarity coefficient of the statistical characteristics between adjacent pulse interval segments and the variation coefficient of the pulse intervals in the same pulse interval segment.

[0139] In the embodiments of the present application, the similarity coefficient of the statistical characteristics between adjacent pulse interval segments is calculated. The difference between the values of each statistical characteristic of the same type is calculated, and all the difference values are integrated (for example, the square root of the sum of the squares). For example, the average interval of segment 1 is 4 seconds, and the standard deviation is 1.414. The average interval of segment 2 is 5 seconds, and the standard deviation is 1.732. The difference is 1 and 0.318, the sum of the squares is approximately 1.101, and the similarity coefficient is approximately 1.05 after taking the square root. The variation coefficient of the same segment is calculated by dividing the interval standard deviation by the average interval. For example, the standard deviation of a segment is 1.414, and the average interval is 4 seconds. The variation coefficient is 0.353.

[0140] In the embodiments of the present application, the similarity coefficient of the statistical characteristics between adjacent pulse interval segments is calculated. The difference between the values of each statistical characteristic of the same type is calculated, and all the difference values are integrated (for example, the square root of the sum of the squares). For example, the average interval of segment 1 is 4 seconds, and the standard deviation is 1.414. The average interval of segment 2 is 5 seconds, and the standard deviation is 1.732. The difference is 1 and 0.318, the sum of the squares is approximately 1.101, and the similarity coefficient is approximately 1.05 after taking the square root. The variation coefficient of the same segment is calculated by dividing the interval standard deviation by the average interval. For example, the standard deviation of a segment is 1.414, and the average interval is 4 seconds. The variation coefficient is 0.353.

[0141] Step 133, associating the statistical characteristics, the similarity coefficient, and the variation coefficient to form a multi-dimensional feature vector.

[0142] In the embodiments of the present application, the similarity coefficient of the statistical characteristics between adjacent pulse interval segments is calculated. The difference between the values of each statistical characteristic of the same type is calculated, and all the difference values are integrated (for example, the square root of the sum of the squares). For example, the average interval of segment 1 is 4 seconds, and the standard deviation is 1.414. The average interval of segment 2 is 5 seconds, and the standard deviation is 1.732. The difference is 1 and 0.318, the sum of the squares is approximately 1.101, and the similarity coefficient is approximately 1.05 after taking the square root. The variation coefficient of the same segment is calculated by dividing the interval standard deviation by the average interval. For example, the standard deviation of a segment is 1.414, and the average interval is 4 seconds. The variation coefficient is 0.353.

[0143] In the embodiments of the present application, the statistical features obtained in the collecting step 131, the similarity coefficient obtained in the step 132 and the variation coefficient are arranged in a preset order to form a multi-dimensional feature vector. For example, the statistical features of a certain segment are [4, 1.414, 2, 6], the similarity coefficient is 1.05, and the variation coefficient is 0.353. After combination, the multi-dimensional feature vector is [4, 1.414, 2, 6, 1.05, 0.353].

[0144] The present application provides the following specific examples: a space radar system processes two adjacent pulse interval segments, segment 1 contains pulse intervals of 1 second, 3 seconds, 3 seconds and 5 seconds, and the statistical features are calculated in the step 131: the average interval is (1+3+3+5) ÷ 4 = 3 seconds, the standard deviation is 1.414, the minimum value is 1 second, and the maximum value is 5 seconds. The histogram distribution is 1 second, 3 seconds twice, and 5 seconds once. In the step 132, the similarity coefficient is calculated with segment 2 (average interval of 4 seconds, standard deviation of 2). The difference is 1 second and 0.586, and the square sum is 1.343. The square root is about 1.16. At the same time, the variation coefficient of segment 1 is calculated as 1.414 ÷ 3 ≈ 0.471. In the step 133, the statistical features [3, 1.414, 1, 5], the similarity coefficient 1.16 and the variation coefficient 0.471 are combined in order to form a multi-dimensional feature vector [3, 1.414, 1, 5, 1.16, 0.471].

[0145] By performing the steps 131-133, the embodiments of the present application extract key data reflecting the basic characteristics of the segment by performing feature statistics on the pulse interval segment, lay a foundation for subsequent analysis; calculate the similarity coefficient and the variation coefficient to supplement the correlation information of adjacent segments and the dispersion degree information of the same segment, enrich the feature dimension; associate these data to form a multi-dimensional feature vector, comprehensively integrate multi-aspect feature information, make the subsequent space radar jamming classification analysis more targeted and comprehensive, and help to improve the classification accuracy.

[0146] In a possible embodiment, the step 132, calculating the similarity coefficient of the statistical features between adjacent pulse interval segments, and the variation coefficient of the pulse intervals in the same pulse interval segment, comprises:

[0147] c1, arranging two pulse interval segments in the pulse interval segment in time sequence and connected in position before and after as adjacent segments.

[0148] Among them, the pulse interval segment is a continuous pulse interval data part divided by time, the time sequence is the arrangement order of the segment from early to late according to the time of appearance, the adjacent segment is two pulse interval segments connected in position before and after in this arrangement, used for analyzing the correlation between segments.

[0149] In the embodiments of the present application, the time sequence of all pulse interval segments is determined, and the two segments adjacent in position in the sequence are determined as adjacent segments, for example, in three segments A, B and C arranged in time sequence, A and B are adjacent segments, and B and C are adjacent segments.

[0150] c2. For each type, the numerical difference value of the statistical feature of the same type in adjacent segments is calculated respectively.

[0151] wherein, the type refers to the category of the statistical feature, the adjacent segments are the two segments determined in step c1, the statistical feature of the same type is the feature data of the same category in the two adjacent segments, and the numerical difference value is the difference between the numerical values of the statistical features of the same type in the two segments, which is used to reflect the difference degree of the features.

[0152] In the embodiments of the present application, all statistical feature types of adjacent segments are listed, and for each type, the statistical feature value of the latter segment is subtracted from the corresponding value of the former segment to obtain the numerical difference value of the type, for example, for adjacent segments A and B, the average interval difference value is 5-4=1 second, and the standard deviation difference value is 1.732-1.414=0.318.

[0153] c3. The numerical difference values of all types of statistical features are integrated to obtain the similarity coefficient between adjacent segments.

[0154] wherein, the numerical difference values of all types of statistical features are the feature difference value set obtained in step c2, the integration processing is a comprehensive calculation process of the difference values, and the similarity coefficient is the value after the integration processing, which is used to measure the similarity degree of the statistical features of adjacent segments, and the smaller the value is, the more similar it is.

[0155] In the embodiments of the present application, all numerical difference values obtained in step c2 are collected, and the similarity coefficient is obtained by processing through a preset method, for example, the difference values 1 and 0.318, the square sum is 1.101, and the square root is about 1.05.

[0156] c4. The average size and dispersion degree value of all pulse interval values contained in the same pulse interval segment are calculated, and the coefficient of variation in the same pulse interval segment is calculated based on the average size and dispersion degree value.

[0157] wherein, the same pulse interval segment is a single pulse interval segment, the all pulse interval values are all pulse interval time data contained in the segment, the average size is the average value of the numerical values, the dispersion degree value is the standard deviation reflecting the dispersion of the numerical values, and the coefficient of variation is the value obtained by dividing the dispersion degree value by the average size, which is used to measure the dispersion degree of the pulse intervals in the segment.

[0158] In the embodiments of the present application, all pulse interval values of the same pulse interval segment are extracted, the average size (the sum of all values divided by the number) and the dispersion value (standard deviation) are calculated, and then the coefficient of variation is obtained by dividing the dispersion value by the average size. For example, the values of a certain segment are 2 seconds, 3 seconds, 5 seconds, 4 seconds, and 6 seconds, the average size is (2+3+5+4+6) ÷ 5 = 4 seconds, the dispersion value is 1.414, and the coefficient of variation is 0.353.

[0159] The present application provides the following specific examples: the pulse interval segments of a space radar system are sequentially arranged as segment 1, segment 2, and segment 3, the c1 step determines that segment 1 and segment 2, and segment 2 and segment 3 are adjacent segments; in the c2 step, the average interval difference value of segment 1 and segment 2 is calculated as 4-3 = 1 second, and the standard deviation difference value is 2-1.414 = 0.586; in the c3 step, the square sum of the two difference values is calculated as 1.343, and the square root is about 1.16; in the c4 step, the pulse interval values of segment 1 are 1 second, 3 seconds, 3 seconds, and 5 seconds, the average size is calculated as (1+3+3+5) ÷ 4 = 3 seconds, the dispersion value is calculated as 1.414, and finally the coefficient of variation is obtained by dividing the dispersion value by the average size, which is about 0.471.

[0160] By performing c1-c4, the embodiments of the present application determine the correlation analysis object by determining adjacent segments, quantify the feature difference between segments by calculating the difference value, the similarity coefficient obtained by integration directly reflects the similarity degree of the segments, and the coefficient of variation measures the dispersion within the same segment. These steps collectively enrich the feature information, provide comprehensive data support for the construction of the subsequent multi-dimensional feature vector, and help improve the accuracy of space radar jamming classification.

[0161] In a possible embodiment, S11, the spacecraft-borne radar or the ground deep space exploration radar collects a pulse description word data stream to be processed, and pulse data segments with significant space jamming features are screened from the pulse description word data stream, including:

[0162] Step 111, extracting the repetition interval of pulse jamming from the pulse description word data stream.

[0163] The pulse description word data stream is a data set containing various information of the pulse signal, the repetition interval of pulse jamming is the time interval at which the pulse jamming signal repeatedly appears, that is, the time difference between adjacent two pulse jamming signals, and the interval is extracted to provide basic data for subsequent analysis of the variation rule.

[0164] In the embodiment of the present application, the signal data of the pulse interference is screened out from the pulse description word data stream, the occurrence time of each pulse interference is recorded, and the difference between the occurrence times of two adjacent pulse interferences is calculated, which is the extracted repetition interval. For example, the occurrence times of the pulse interference are 2 seconds, 5 seconds, 8 seconds and 12 seconds, the difference is calculated as 5-2=3 seconds, 8-5=3 seconds and 12-8=4 seconds, and the repetition intervals are 3 seconds, 3 seconds and 4 seconds.

[0165] Step 112, identify the periodicity and randomness type of the repetition interval.

[0166] The periodicity type of the repetition interval is that the interval presents a fixed regular change mode, and the randomness type is that the interval changes randomly and the value is random. The two types are identified by analyzing the change rule of the interval to determine whether it belongs to periodicity or randomness.

[0167] In the embodiment of the present application, the repetition intervals obtained in step 111 are collected, and whether the value changes regularly is analyzed. If the interval repeats in a fixed period, it is identified as a periodic type. If the interval has no fixed mode, it is identified as a randomness type. For example, the intervals of 3 seconds, 3 seconds, 4 seconds, 3 seconds, 3 seconds and 4 seconds are identified as periodic because they repeat the same pattern, and the intervals of 3 seconds, 5 seconds, 2 seconds and 6 seconds are identified as random because they have no regularity.

[0168] Step 113, based on the periodicity, calculate the proportion of the randomness type in all types of repetition intervals, and according to the proportion, screen out the pulse data segment with significant periodicity characteristics or significant randomness characteristics.

[0169] The proportion of the randomness type is the proportion of the number of random intervals in the total interval, the significant periodicity characteristic is the characteristic that the proportion of the periodic interval is high, the significant randomness characteristic is the characteristic that the proportion of the random interval is high, and the screening is to retain the data segment that meets the condition according to the comparison of the proportion with the threshold value.

[0170] In the embodiment of the present application, the total number of repetition intervals and the number of random intervals are counted, the randomness proportion is calculated, and after setting the threshold value, if the proportion is lower than the threshold value, the significant periodicity data segment is screened out, and if the proportion is higher than the threshold value, the significant randomness data segment is screened out. For example, 2 of the 10 intervals are random, the proportion is 2÷10=20%, and if the threshold value is 20%, it is a significant periodicity; 8 of them are random, the proportion is 80%, and it is a significant randomness.

[0171] The application provides the following specific examples: when a space radar system processes a pulse description word data stream, step 111 extracts the repetition intervals of pulse interference from the data stream to obtain 3 seconds, 3 seconds, 4 seconds, 3 seconds, 3 seconds, 4 seconds, 5 seconds, 3 seconds, 3 seconds and 4 seconds; step 112 analyzes and finds that 8 intervals conform to the periodic pattern of 3 seconds, 3 seconds and 4 seconds, and 2 intervals are 5 seconds; step 113 counts the total interval number as 10, calculates the randomness ratio as 2÷10=20%, sets the significant periodicity threshold as 20%, and therefore the data segment is screened as a pulse data segment with a significant periodicity feature.

[0172] By performing steps 111-113, the embodiment of the application provides a basis for analyzing interference rules by extracting the repetition intervals of pulse interference; can distinguish the change mode of interference signals by identifying periodicity and randomness types; can focus on key data by screening significant feature data segments based on the ratio, reduce irrelevant interference, provide accurate analysis objects for subsequent space radar interference classification, and improve the effectiveness of classification processing.

[0173] Figure 3 A structure diagram of a radar interference classification system based on an attention mechanism provided by the embodiment of the application is shown in FIG. 1. Figure 3 As shown in FIG. 1, the system comprises:

[0174] The acquisition module 31 is configured to acquire a pulse description word data stream to be processed collected by a spacecraft-borne radar or a ground deep space exploration radar, and screen a pulse data segment with a significant space interference feature from the pulse description word data stream.

[0175] The division module 32 is configured to perform time sequence positioning on the pulse data segment by using an ephemeris timestamp marking method, divide the pulse data segment after time sequence positioning into a plurality of pulse interval segments by adapting a preset pulse threshold of a space target motion cycle, and cause each pulse interval segment to correspond to one or more pulse intervals.

[0176] The processing module 33 is configured to perform a multi-stage processing of the space scene adaptation on the plurality of pulse interval segments to obtain a multi-dimensional feature vector.

[0177] The output module 34 is configured to input the multi-dimensional feature vector into a pre-trained space radar interference classification neural network model, the model uses a convolutional neural network to extract local space-time features, uses a multi-head attention mechanism to capture long-range dependency relationships between features and enhance key features, and outputs a probability distribution of a space radar interference type corresponding to the multi-dimensional feature vector through a Softmax classification layer.

[0178] The determination module 35 is configured to determine a space radar interference classification result according to a maximum probability in the probability distribution.

[0179] The attention mechanism-based radar interference classification system of the embodiments of the present application is used to implement the aforementioned attention mechanism-based radar interference classification method, and therefore the specific embodiments of the attention mechanism-based radar interference classification system can be seen from the aforementioned embodiment part of the attention mechanism-based radar interference classification method, and the specific embodiments can be referred to the description of the corresponding embodiment part, which will not be described here.

[0180] The present application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of the aforementioned attention mechanism-based radar interference classification method when executing the computer program.

[0181] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the aforementioned attention mechanism-based radar interference classification method.

[0182] In an exemplary embodiment, the aforementioned computer readable storage medium can include, but is not limited to, a U disk, a read-only memory, a random access memory, a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0183] The embodiments of the present application also provide a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the steps in the aforementioned attention mechanism-based radar interference classification method.

[0184] The skilled person can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in a general manner in the foregoing description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0185] The above describes in detail the attention mechanism-based radar interference classification method, system, electronic device and storage medium provided by the present application. The principles and implementation manners of the present application are described by using specific examples in this paper, and the above description of the examples is only used to help understand the method and core idea of the present application. It should be pointed out that for ordinary skilled person in the technical field, without departing from the principles of the present application, some improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A radar jamming classification method based on an attention mechanism, characterized in that, include: Acquire pulse descriptor data streams to be processed from spacecraft-borne radar or ground-based deep space exploration radar, and filter out pulse data segments with significant characteristics of space interference from the pulse descriptor data streams; The pulse data segment is time-series located using the ephemeris timestamp marking method. The pulse data segment after time-series location is divided into multiple pulse interval segments by a preset pulse threshold that adapts to the motion cycle of the space target. Each pulse interval segment corresponds to one or more pulse intervals. The multiple pulse interval segments are subjected to multi-level processing for aerospace scenario adaptation to obtain multi-dimensional feature vectors; The multidimensional feature vector is input into a pre-trained aerospace radar interference classification neural network model. The model uses a convolutional neural network to extract local spatiotemporal features and uses a multi-head attention mechanism to capture long-range dependencies between features and enhance key features. The probability distribution of the aerospace radar interference type corresponding to the multidimensional feature vector is output through a Softmax classification layer. The classification result of aerospace radar interference is determined based on the maximum probability in the probability distribution; The multidimensional feature vector is input into a pre-trained aerospace radar interference classification neural network model. The model uses a convolutional neural network to extract local spatiotemporal features and employs a multi-head attention mechanism to capture long-range dependencies between features and enhance key features. The model outputs the probability distribution of the aerospace radar interference type corresponding to the multidimensional feature vector through a Softmax classification layer, including: The convolutional layer of the convolutional neural network based on the pre-trained interference classification neural network model extracts local features from the multidimensional feature vector through a sliding window to obtain a local feature sequence. After introducing nonlinearity through the ReLU activation function of the convolutional neural network, the feature dimension is compressed by a max pooling layer of a preset size, and the dimension of the local feature sequence is adjusted to retain key local features, thus obtaining the adjusted local feature sequence. The local feature sequence is mapped to a first feature matrix of a preset dimension through the fully connected layer of the convolutional neural network, so that the first feature matrix is ​​adapted to the input requirements of the multi-head attention mechanism. Generate a position encoding vector that matches the length of the first feature matrix, and add the position encoding vector to the first feature matrix to obtain the second feature matrix; The second feature matrix is ​​divided into multiple feature subsequences, and each feature subsequence corresponds to an attention head; Calculate the correlation value between each attention head and the corresponding feature subsequence; Based on the correlation degree value, the feature subsequences are weighted and fused to obtain multiple attention features, and the multiple attention features are concatenated into a fused feature sequence. Based on the Softmax classification layer in the interference classification neural network model, the probability distribution of the aerospace radar interference type corresponding to the multidimensional feature vector is obtained by numerically converting the matching degree between the fused feature sequence and various types of aerospace radar interference. The multi-level processing of the multiple pulse interval segments for aerospace scenario adaptation yields a multi-dimensional feature vector, including: The characteristic statistics of each pulse interval segment are performed to obtain the corresponding statistical features, including at least one of the following: average interval, standard deviation of interval, minimum interval value, maximum interval value, and interval histogram distribution. Calculate the similarity coefficient of statistical characteristics between adjacent pulse interval segments, and the coefficient of variation of pulse intervals within the same pulse interval segment; The statistical features, the similarity coefficient, and the coefficient of variation are correlated to form a multidimensional feature vector.

2. The method according to claim 1, characterized in that, The calculation of the correlation value between each attention head and the corresponding feature subsequence includes: The feature subsequence is converted into a first feature component and a second feature component, wherein the first feature component and the second feature component have the same feature dimension. Calculate the correlation strength values ​​between each feature component in the first feature component and each feature component in the second feature component, and construct a numerical set based on each correlation strength value; Each correlation strength value in the numerical set is adjusted so that the adjusted correlation strength value is within a preset range. Based on the adjusted correlation strength value, the correlation values ​​between features at each position in the feature subsequence are determined. Based on the preset weight parameters corresponding to each attention head, the associated values ​​are weighted and integrated to obtain the degree of association between each attention head and the corresponding feature subsequence.

3. The method according to claim 2, characterized in that, The step of adjusting each correlation strength value in the numerical set so that the adjusted correlation strength value is within a preset range, and determining the correlation value between features at each position in the feature subsequence based on the adjusted correlation strength value, includes: Set an upper and lower limit value for a preset interval, and judge each correlation strength value in the set of values; If the judgment result indicates that the correlation strength value is greater than the upper limit value, then the upper limit value is used as the correlation strength value. If the judgment result indicates that the correlation strength value is less than the lower limit value, then the lower limit value is used as the correlation strength value. If the judgment result indicates that the correlation strength value is between the upper limit value and the lower limit value, then the correlation strength value is kept unchanged to obtain an adjusted value. The adjusted value is positively correlated with the correlation relationship between each feature in the feature subsequence. Based on all the adjusted values, the correlation values ​​between features in the feature subsequence are determined.

4. The method according to claim 1, characterized in that, The calculation of the similarity coefficient of statistical characteristics between adjacent pulse interval segments and the coefficient of variation of pulse intervals within the same pulse interval segment includes: Two pulse interval segments arranged in chronological order and whose positions are consecutive are considered as adjacent segments. For each type, the corresponding numerical difference is calculated for the statistical features of the same type in the adjacent segments; The numerical differences corresponding to all types of statistical features are integrated to obtain the similarity coefficient between the adjacent segments; Extract all pulse interval values ​​contained in the same pulse interval segment, calculate the average size and dispersion value of all pulse interval values, and calculate the coefficient of variation within the same pulse interval segment based on the average size and dispersion value.

5. The method according to claim 1, characterized in that, The step of filtering pulse data segments with significant characteristics of aerospace interference from the pulse descriptor data stream includes: Extract the repetition interval of the pulse interference from the pulse descriptor data stream; Identify the periodicity and randomness type of the repeating interval; Based on the periodicity, the proportion of the randomness type in all types of the repeating interval is calculated, and pulse data segments with significant periodicity or significant randomness are selected according to the proportion.

6. A radar jamming classification system based on an attention mechanism, characterized in that, For performing the radar jamming classification method based on the attention mechanism as described in claim 1, including: The acquisition module is used to acquire the pulse descriptor data stream to be processed collected by spacecraft-borne radar or ground-based deep space exploration radar, and to filter out pulse data segments with significant characteristics of space interference from the pulse descriptor data stream. The segmentation module is used to perform time-series positioning of the pulse data segment using the ephemeris timestamp marking method. By using a preset pulse threshold that adapts to the motion cycle of the space target, the time-series positioned pulse data segment is divided into multiple pulse interval segments, and each pulse interval segment corresponds to one or more pulse intervals. The processing module is used to perform multi-level processing on the multiple pulse interval segments for aerospace scenario adaptation to obtain multi-dimensional feature vectors; The output module is used to input the multidimensional feature vector into a pre-trained aerospace radar interference classification neural network model. The model uses a convolutional neural network to extract local spatiotemporal features and uses a multi-head attention mechanism to capture long-range dependencies between features and enhance key features. The model outputs the probability distribution of the aerospace radar interference type corresponding to the multidimensional feature vector through a Softmax classification layer. The determination module is used to determine the aerospace radar interference classification result based on the maximum probability in the probability distribution.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the attention-based radar jamming classification method as described in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the radar jamming classification method based on an attention mechanism as described in any one of claims 1 to 5.

Images