Radar signal sorting method based on improved fuzzy adaptive resonance network

Abstract

The present invention belongs to the technical field of signal sorting, and relates to a radar signal sorting method based on an improved fuzzy adaptive resonance network. It includes: constructing a radar mixed pulse sequence, constructing a feature set based on feature parameters and normalizing it; constructing an improved fuzzy adaptive resonance network; inputting the normalized feature set samples into the improved fuzzy adaptive resonance network for processing, specifically including visual evaluation of clustering trends, selecting winning neurons and calculating the matching degree between the input samples and the winning neurons; judging whether the matching degree exceeds the warning threshold, and cyclically updating the weight vector, sending the obtained category matrix to the category merging module for category merging, and outputting the final sample classification result; and then performing index evaluation; the method enhances the network memory, makes it insensitive to the input order, adapts to different input modes, and avoids the phenomenon of class appreciation; when the number of categories is unknown, it realizes accurate determination of the number of categories, greatly improves the sorting accuracy, and runs faster and more stable.

Description

Technical Field

[0001] The invention belongs to the technical field of data representation and signal sorting, and in particular relates to a radar signal sorting method based on an improved fuzzy adaptive resonance network. Background Art

[0002] In the actual electromagnetic environment, the radiation source signals intercepted by the reconnaissance receiver are usually intertwined pulse sequences. In order to effectively grasp the information such as the types, properties and threat levels of different radiation sources, it is necessary to accurately extract all the pulses contained in them. Signal sorting is the key technology to complete this task. It is an important link in the radar reconnaissance signal processing process, which is related to the subsequent enemy and friendly situation judgment and combat decision-making. Its performance has also become an important indicator for evaluating whether the radar reconnaissance system can adapt to the electromagnetic environment. Therefore, it is extremely important to achieve efficient and accurate signal sorting.

[0003] Signal sorting technology can be roughly divided into three categories according to the different characteristic parameters used: sorting algorithms based on pulse repetition interval PRI, sorting algorithms based on intra-pulse features, and sorting algorithms based on multiple parameters. The algorithm based on pulse repetition interval extracts the pulse sequence based on the PRI characteristics of the signal. This type of method plays an important role in the field of signal sorting, but with the advancement of radar technology, the changes in PRI have become more and more diverse, which has caused certain obstacles to the sorting method that only uses PRI parameters. The sorting algorithm based on intra-pulse features first extracts the intra-pulse characteristic parameters unique to the pulse signal from the transform domain, and then sends these features to the classifier for sorting and identification. This type of algorithm has a large amount of calculation, and the extracted features are not generalizable and robust. In the case of low signal-to-noise ratio, the intra-pulse features are unstable and have weak adaptability. Sorting based on multiple parameters refers to the technology that uses multiple characteristic parameters such as intra-pulse or inter-pulse of radar pulse signals for classification, which mainly uses clustering algorithms as signal classifiers. Cluster sorting refers to using the characteristic parameters of radar pulse signals as the input of the algorithm, and by calculating the differences between the characteristic parameters, all pulses belonging to the same radiation source are classified into the same category as much as possible. This type of algorithm does not require the allocation of sample data for training and learning in advance, and is an unsupervised classification. This is of great value for radar signal sorting that lacks prior knowledge, and is also a major reason why clustering algorithms are widely used.

[0004] With the development of technology, artificial neural networks have been applied to the field of radar signal sorting. The signal sorting method based on neural network is essentially a model clustering method, which has the ability of intelligent learning and parallel processing, and can adapt well to complex, unknown, dynamic and changeable electromagnetic environments. Fuzzy Adaptive Resonance Theory (Fuzzy ART) is a commonly used sorting method based on neural networks. The stability and plasticity of the network enable it to self-organize and adjust its structure when facing unknown radiation source signals. It does not require manual pre-setting, and the network can make good use of existing clustering information to improve the efficiency of training and learning. However, the original network still has two shortcomings: it is sensitive to the order of sample input and is affected by fixed thresholds and sparse points to produce class value-added problems.

[0005] In order to solve the problem that traditional algorithms need prior information and cannot perform intelligent sorting in unknown and complex electromagnetic environments, the present invention proposes a sorting algorithm based on an improved fuzzy adaptive resonance network, namely a merged fuzzy ART network (Merged Fuzzy ART based on VAT and Adaptive vigilance threshold, VAMFA) based on the Fuzzy ART network, which further improves the sorting performance of the original network. The simulation results show that on radar feature sets with severe overlap and lack of prior information, the proposed VAMFA network model can achieve accurate determination of the number of classes when the number of classes is unknown, and greatly improves the sorting accuracy compared to the original network and several latest improved networks, and has a faster running speed and stronger stability. Summary of the invention

[0006] The purpose of the present invention is to address the shortcomings of existing sorting algorithms in unknown complex electromagnetic environments, such as the need for prior information and the sensitivity of fuzzy adaptive resonance networks to the order of sample input and the problem of class value-added caused by fixed thresholds and sparse points. A radar signal sorting method based on an improved fuzzy adaptive resonance network is proposed. The method reorders samples based on visual evaluation of clustering trends, constructs an adaptive warning threshold, considers the influence of current input and uses classified results as feedback, effectively merges redundant categories caused by sparse sample points, and reduces the influence of sparse points on classification results.

[0007] In order to achieve the above purpose, the following technical solutions are adopted.

[0008] The radar signal sorting method comprises the following steps:

[0009] S1, constructing a radar mixed pulse sequence, building a feature set based on characteristic parameters and normalizing it to obtain a normalized feature set;

[0010] S1 specifically includes the following sub-steps:

[0011] S11, construct radar mixed pulse sequence;

[0012] There are M types of radar mixed pulse sequences; each radar includes a number of signal pulses and the dimension of each signal pulse is D, that is, the signal pulse is represented by D characteristic parameters;

[0013] S12, collecting signal pulses to be sorted from the radar mixed pulse sequence constructed in S11, and recording the total number of signal pulses to be sorted as N;

[0014] S13, forming a feature set from the feature parameters of each signal pulse to be sorted, and performing normalization processing on the feature set, that is, changing each dimensional feature in the feature set to the same scale range to obtain a normalized feature set;

[0015] The total number of feature sets is the same as the total number of signal pulses to be sorted, which is N. The feature parameters of each signal pulse in the feature set can be regarded as N samples.

[0016] The normalization process adopts maximum and minimum normalization;

[0017] S2, construct an improved fuzzy adaptive resonance network;

[0018] The improved fuzzy adaptive resonance network includes a preprocessing module, an input layer, a membership calculation layer, a category representation layer and a category merging module;

[0019] The input layer, membership calculation layer and category representation layer are denoted as F0, F1 and F2 respectively;

[0020] Among them, the F1 layer and the F2 layer are connected by a weight vector W with a dimension of 2D;

[0021] S3, inputting the samples in the normalized feature set into the improved fuzzy adaptive resonance network constructed in S2 for processing, and outputting the sample classification results, specifically including the following sub-steps:

[0022] S31, input all N samples in the normalized feature set into the preprocessing module for preprocessing, that is, reorder the samples according to the similarity between them to obtain reordered input samples; since the above operation can see the cluster trend from the reordered input samples, it is also called cluster trend visual assessment technology, specifically:

[0023] Initialize the index matrix and set it to empty; calculate the distance dissimilarity matrix between samples, take one of the two samples with the farthest distance as the starting point, and add its corresponding position in the feature set to the index matrix; according to the minimum distance to any sample in the current index matrix, iterate according to the single-link hierarchical clustering and add the corresponding positions of the remaining samples in the feature set until all samples are traversed to obtain the final index matrix; rearrange the order of samples according to the index matrix to obtain the reordered input samples;

[0024] S32, set the iteration loop variable n, and set the initial loop variable to n=1; the category matrix of the network output is recorded as S, and the initial category matrix is ​​set to S=[ ], where the rows represent the numbers of the input samples contained in different categories, and the columns represent the number of categories; when the weight vector is initialized, the values ​​of all elements are set to 1;

[0025] S33, sending the nth input sample reordered in S31 to F0, performing a complement coding operation and connecting it in series with the input sample to obtain an expanded input sample;

[0026] The expanded input sample, i.e., the output of F0;

[0027] The complementary coding operation specifically comprises: obtaining a complementary code for each dimension of the sample to obtain a sample complementary code; and concatenating the sample complementary code with the current sample to form an extended input sample.

[0028] The dimension of the expanded input sample is 2D;

[0029] S34, sending the input sample expanded by S32 and the weight vector from top to bottom of the F2 layer to the F1 layer, performing a mode selection operation, that is, calculating the membership degree of the input sample and the neuron to activate the corresponding neuron of the F2 layer;

[0030] The mode selection operation is as follows: take the minimum value of the elements in each dimension of the input sample and the weight vector of each connected neuron, calculate the 1 norm of the obtained vector as the numerator; calculate the 1 norm of each weight vector and add it to the selection factor as the denominator; the ratio of the numerator to the denominator is the membership degree; select the F2 layer neuron corresponding to the maximum membership degree as the winning neuron; where the selection factor is recorded as , with a value range of 0 to 1; the weight vector corresponding to the winning neuron is recorded as W J ;

[0031] S35, the weight vector W corresponding to the input sample expanded by S32 and the winning neuron selected by S33 J Send it to the F2 layer to perform pattern matching operation, that is, calculate the matching degree between the input sample and the winning neuron;

[0032] The matching degree is as follows: Take the input sample and the weight vector W J The minimum value of the elements in each dimension, the 1 norm of the obtained vector is calculated as the numerator; the 1 norm of the input sample is calculated as the denominator; the ratio of the numerator to the denominator is the matching degree;

[0033] S36: Determine whether the matching degree calculated in S35 exceeds a warning threshold, where the warning threshold is ; If greater than or equal to , the reset node sends a resonance signal, the network starts learning, adds the nth input sample to the input samples of the corresponding category in the S matrix, updates the weight vector, and jumps to S36, where the reset node is recorded as R; if it is less than , the network mismatch is reset, R sends a reset signal, and the network re-searches for neurons that meet the matching degree exceeding the warning threshold; if the matching degree of all neurons in the F2 layer and the input samples cannot exceed the warning threshold, the nth input sample is added to the F2 layer as a new neuron, the S matrix adds a column, and the input sample is used as a new weight vector;

[0034] The calculation of the warning threshold is as follows:

[0035] S361, calculate the fuzzy distance between the weight vector WJ corresponding to the winning node and the input sample in the kth dimension as the reference distance, denoted as , where k ranges from [1,2D]; the fuzzy distance is equal to 1 minus the inverse exponential of the absolute value of the weight vector WJ minus the value of the input sample on the kth dimension; where the fuzzy factor is denoted by m and ranges from 0 to 20;

[0036] S362, calculate the fuzzy distance between the weight vector WJ corresponding to the winning node and all classified input samples belonging to the winning node in the kth dimension as the weighted distance, denoted as , where the weighted distance is equal to the fuzzy distance between each input sample and the weight vector WJ multiplied by the value of the input sample in the kth dimension, divided by the sum of the elements of all input samples in the kth dimension;

[0037] S363, comparing the reference distance obtained in S361 with the weighted distance obtained in S362, and taking the smaller value between the two as the value of the warning threshold in the kth dimension;

[0038] S364: Select the maximum value of all elements of the warning threshold as the final warning threshold for matching judgment ;

[0039] S37, update the weight vector, specifically: take the minimum value of the input sample and the element in each dimension of the weight vector, and multiply the obtained vector by the learning factor, where the learning factor is recorded as , the value range is [0,1]; calculate the weight vector and 1- The result after multiplication; add the two results obtained by the above calculation as a new weight vector;

[0040] S38, determine whether n is greater than the total number of signals to be sorted N, if not greater than N, update n: n=n+1, jump to S33; if greater than N, the iterative loop ends, jump to S39;

[0041] S39, sending the category matrix C matrix outputted from S36 to the category merging module, performing the category merging operation, and outputting the final sample classification result;

[0042] The category merging operation is as follows:

[0043] S391, counting the number of input samples contained in each category in the category matrix C, and recording the number of categories in the S matrix as P;

[0044] S392, set the iteration loop variable p, and set the initial loop variable to p=1;

[0045] S393, determine whether the number of samples in the pth category exceeds the threshold. If not, take the mean of the input sample set corresponding to all categories in the F2 layer as the center of each category, merge the category number and the category center, and obtain a new category matrix C1; if it exceeds, jump to S394; wherein the threshold is equal to N divided by P divided by 2;

[0046] S394, determine whether p is greater than the number of categories P. If not, update p: p=p+1, and jump to S393; if greater than P, use the number of columns of the new category matrix C1 as the number of categories after merging P1, and the iterative loop ends;

[0047] S395, set the iteration loop variable p1, and set the initial loop variable to p1=1;

[0048] S396. When p1 is equal to 1, the mean of the input sample set corresponding to all categories in the category matrix C1 is taken as the center of each category, and the distance between the category centers is calculated; when p1 is greater than 1, the mean of the input sample set corresponding to all categories in the category matrix C2 is taken as the center of each category, and the distance between the category centers is calculated;

[0049] S397, determine whether the distance between the two category centers exceeds the threshold. If not, merge the corresponding two category numbers and category centers to obtain a new category matrix C2; if exceeded, jump to S398, where the threshold is equal to the sum of the distances between all N sample points divided by N to the power of a and N-1 to the power of a; where the value range of a is 0 to 1;

[0050] S398, determine whether p1 is greater than the number of categories P1. If not, update p1: p1=p1+1, and jump to S396; if greater than P1, output the new category matrix C2 and the category center corresponding to each category as the final sample classification result, and the iterative loop ends;

[0051] S4, evaluating the sample classification results obtained in S3 based on the evaluation indicators;

[0052] Among them, the evaluation indicators include the number of clusters, cluster quality, centroid index and running time;

[0053] So far, from S1 to S4, the radar signal sorting method based on the improved fuzzy adaptive resonance network has been completed.

[0054] Beneficial Effects

[0055] Compared with the existing signal sorting methods, the radar signal sorting method based on the improved fuzzy adaptive resonance network of the present invention has the following beneficial effects:

[0056] 1. The method uses clustering trend visual assessment VAT technology as a preprocessor and reorders the radar pulse sequence based on similarity measurement, making the sample distribution more regular and reducing the impact of sample input order on network stability;

[0057] 2. The method constructs an adaptive warning threshold, which takes into account both the distance information between the current input and the center of the category to which the network defines it, and the distance information between the previous input samples in the category and the category center, thereby enhancing the memory of the network so that the network can adapt to different input modes and avoid the occurrence of class appreciation phenomenon;

[0058] 3. The method effectively merges the redundant categories caused by sparse sample points, reducing the impact of sparse points on the classification results;

[0059] 4. Compared with the original fuzzy adaptive resonance network and several latest improved networks, the method can accurately determine the number of categories when the number of categories is unknown, while greatly improving the sorting accuracy. It also has a faster running speed and stronger stability, and realizes efficient and accurate signal sorting in unknown and complex electromagnetic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A three-dimensional characteristic distribution diagram of radar signals of the radar signal sorting method based on an improved fuzzy adaptive resonance network of the present invention;

[0061] Figure 2 A schematic diagram of the network structure of a radar signal sorting method based on an improved fuzzy adaptive resonance network according to the present invention;

[0062] Figure 3 It is a flow chart of a radar signal sorting method based on an improved fuzzy adaptive resonance network of the present invention;

[0063] Figure 4 It is a schematic diagram of the network sorting performance of the radar signal sorting method based on the improved fuzzy adaptive resonance network under different warning thresholds of the present invention;

[0064] Figure 5 This is the signal sorting result of the radar signal sorting method based on the improved fuzzy adaptive resonance network of the present invention. DETAILED DESCRIPTION

[0065] The radar signal sorting method based on the improved fuzzy adaptive resonance network of the present invention is further illustrated and described in detail below in conjunction with the accompanying drawings and embodiments.

[0066] Example 1

[0067] This embodiment uses pulse description words (PDW) that can characterize radar signal characteristics for signal sorting. In PDW, PA has the worst stability and is usually not involved in cluster sorting. TOA cannot directly reflect the characteristics of intercepted signals, but is used to extract deeper PRI features to assist signal sorting. Therefore, this embodiment uses DOA, RF and PW three-dimensional features as the input of the algorithm. In order to effectively evaluate the improved fuzzy adaptive resonance network (such as Figure 2 (As shown) In the face of the sorting performance of complex overlapping intercepted pulse sequences, the construction of electromagnetic scenes is particularly important. This embodiment uses 15 groups of mixed sequences generated by a radar signal generator for sorting. The characteristic parameters of different radiation sources in the sequence have different degrees of overlap and the number of pulses is unevenly distributed. Among them, DOA is the relative angle between the radar and the reconnaissance aircraft, which has nothing to do with the radiation source signal itself, and because the radar will not change its azimuth rapidly in a short period of time, the range of variation of the DOA parameter is very small, generally within ±3°. There is a certain range of variation in both PW and RF parameters, where the PW variation range is within ±5µs and the RF variation range is within ±300MHz. The characteristic parameters of each radar signal are specifically shown in Table 1:

[0068] Table 1 Radar signal characteristic parameters

[0069]

[0070] It can be seen from Table 1 above that the characteristic parameters of each group of radar radiation source signals are similar to those of other groups in one or two dimensions. For example, the 8th to 10th groups of radiation sources can be regarded as radiation source signals from different directions. Their pulse width and carrier frequency variation ranges are roughly the same, but there are certain differences in the arrival angles; the 12th to 15th groups of radiation sources can be regarded as radiation source signals from the same direction. Their arrival angles and pulse width variation ranges are roughly the same, but there are differences in carrier frequencies. It can be seen that when constructing the radar signal set used for sorting, the various possibilities of signals intercepted by the receiver from a complex electromagnetic environment are comprehensively considered. The normalized three-dimensional feature distribution of the 15 groups of radar signals is shown in the figure below. Figure 1 shown.

[0071] The radar signal sorting method of this embodiment performs the following steps: Figure 3 As shown, it specifically includes the following sub-steps:

[0072] S1, constructing a radar mixed pulse sequence, building a feature set based on characteristic parameters and normalizing it to obtain a normalized feature set;

[0073] S1 specifically includes the following sub-steps:

[0074] S11, construct radar mixed pulse sequence;

[0075] There are M types of radar mixed pulse sequences; each radar includes a number of signal pulses and each signal pulse has a dimension of D, that is, the signal pulse is represented by D characteristic parameters; in a specific implementation, M is equal to 15 and D is equal to 3;

[0076] S12, collecting the signal pulses to be sorted from the radar mixed pulse sequence constructed in S11, and recording the total number of the signal pulses to be sorted as N; in specific implementation, N is equal to 2000;

[0077] S13, forming a feature set from the feature parameters of each signal pulse to be sorted, and performing normalization processing on the feature set, that is, changing each dimensional feature in the feature set to the same scale range to obtain a normalized feature set;

[0078] The total number of feature sets is the same as the total number of signal pulses to be sorted, which is N. The feature parameters of each signal pulse in the feature set can be regarded as N samples.

[0079] The normalization process adopts maximum and minimum normalization, and the calculation is as follows:

[0080] (1)

[0081] Among them, x represents each data sample; x maxRepresents the maximum value of the radar feature set; x min Minimum value of radar signature set;

[0082] S2, construct an improved fuzzy adaptive resonance network;

[0083] The improved fuzzy adaptive resonance network, such as Figure 2 As shown, it includes a preprocessing module, an input layer, a membership calculation layer, a category representation layer and a category merging module;

[0084] The input layer, membership calculation layer and category representation layer are denoted as F0, F1 and F2 respectively;

[0085] Among them, the F1 layer and the F2 layer are connected by a weight vector W with a dimension of 2D;

[0086] S3, inputting the samples in the normalized feature set into the improved fuzzy adaptive resonance network constructed in S2 for processing, and outputting the sample classification results, specifically including the following sub-steps:

[0087] S31, input all N samples in the normalized feature set into the preprocessing module for preprocessing, that is, re-order them according to the similarity between the samples, perform a visual evaluation of the clustering trend, and obtain the re-ordered input samples, specifically:

[0088] S311, initialize the index matrix and set it to empty; calculate the distance dissimilarity matrix between samples, take one of the two samples with the farthest distance as the starting point, and add its corresponding position in the feature set to the index matrix, and calculate as follows:

[0089] (2)

[0090] in, is the distance dissimilarity matrix; The index after reordering the samples;

[0091] S312, according to the minimum distance to any sample in the current index matrix, iterate according to the single-link hierarchical clustering and add the corresponding positions of the remaining samples in the feature set until all samples are traversed to obtain the final index matrix, which is calculated as follows:

[0092] (3)

[0093] S313, rearrange the order of samples according to the index matrix to obtain rearranged input samples;

[0094] S32, initializing the iteration loop variable n to 1, the category matrix C to be empty, and all elements of the weight vector to be 1;

[0095] S32: the category matrix C is the output of the improved fuzzy adaptive resonance network, and the rows in C represent the numbers of the input samples contained in different categories, and the columns represent the number of categories;

[0096] S33, sending the nth input sample reordered in S31 to the F0 layer to perform a complement coding operation to obtain a sample complement, and concatenating the sample complement with the nth sample to obtain an extended input sample;

[0097] The dimension of the expanded input sample is 2D, which is the output of the F0 layer;

[0098] The complement coding operation is specifically as follows: find the complement code for each dimension of the sample to obtain the sample complement code; concatenate the sample complement code with the current sample to form an extended input sample, which is calculated as follows:

[0099] (5)

[0100] in, is the input sample after complement encoding;

[0101] S34, sending the input sample expanded by S33 and the weight vector from the top to the bottom of the F2 layer to the F1 layer, performing a mode selection operation, calculating the membership of the input sample and the neuron, and selecting the F2 layer neuron corresponding to the maximum membership as the winning neuron;

[0102] The weight vector corresponding to the winning neuron is recorded as W J ;

[0103] The mode selection operation is specifically as follows:

[0104] S341, take the minimum value of the elements in each dimension of the input sample and the weight vector of each connected neuron, calculate the 1 norm of the obtained vector as the numerator; calculate the 1 norm of each weight vector, and add it to the selection factor as the denominator; the ratio of the numerator to the denominator is the membership degree, which is calculated as follows:

[0105] (6)

[0106] in, is the number of pattern nodes in the F2 layer; is the selection function, which represents the similarity between the input pattern and the output neuron; The operator represents the 1-norm of a vector; in practice, the factor The value is 0.1;

[0107] S342, selecting the F2 layer neuron with the largest membership degree as the winning neuron;

[0108] S35, the weight vector W corresponding to the input sample expanded by S33 and the winning neuron selected by S34 J Send it to the F2 layer to perform pattern matching operation, that is, calculate the matching degree between the input sample and the winning neuron;

[0109] The matching degree is as follows: Take the nth input sample and the weight vector W J The minimum value of the elements in each dimension, calculate the 1 norm of the obtained vector as the numerator; calculate the 1 norm of the input sample as the denominator; calculate the ratio of the numerator to the denominator, that is, the matching degree of the input sample with the winning neuron, as follows:

[0110] (7)

[0111] in, is the matching function, and its value range is [0,1];

[0112] S36, determine whether the matching degree calculated by S35 exceeds the warning threshold; if it is greater than or equal to the warning threshold, the reset node sends a resonance signal, the network starts learning, and the nth input sample is added to the input samples contained in the corresponding category in the C matrix, and the weight vector is updated, and jumps to S37; otherwise, if it is less than the warning threshold, the network mismatch is reset, the reset node sends a reset signal, and the network re-searches for neurons that meet the matching degree exceeding the warning threshold; if the matching degree of all neurons in the F2 layer and the input samples cannot exceed the warning threshold, the nth input sample is added to the F2 layer as a new neuron, and the nth input sample is used as a new weight vector and added to the category matrix C to obtain the updated category matrix C;

[0113] The warning threshold in S36 is recorded as , the reset node is denoted as R;

[0114] The calculation of the warning threshold is as follows:

[0115] S361. Calculate the weight vector W corresponding to the winning node J The fuzzy distance between the input sample in the kth dimension is taken as the reference distance, denoted as , where k ranges from [1,2D], and the fuzzy distance is equal to 1 minus the fuzzy factor multiplied by the weight vector W J The inverse exponential of the absolute value of the input sample after subtracting the value on the kth dimension is calculated as follows:

[0116] (8)

[0117] in, is the distance vector; m is the fuzzy factor, and in specific implementation, m is equal to 12;

[0118] S362, calculate the fuzzy distance between the weight vector WJ corresponding to the winning node and all classified input samples belonging to the winning node in the kth dimension as the weighted distance, denoted as , where the weighted distance is equal to the fuzzy distance between each input sample and the weight vector WJ multiplied by the value of the input sample in the kth dimension, divided by the sum of the elements of all input samples in the kth dimension, calculated as follows:

[0119] (9)

[0120] in, is the input sample set corresponding to the winning node; is the distance vector;

[0121] S363, compare the reference distance obtained in S361 and the weighted distance obtained in S362, and take the smaller value between the two as the value of the warning threshold in the kth dimension, which is calculated as follows:

[0122] (10)

[0123] S364: Select the maximum value of all elements of the warning threshold as the final warning threshold for matching judgment , which is calculated as follows:

[0124] (11)

[0125] S37, update the weight vector, specifically: take the minimum value of the input sample and the element in each dimension of the weight vector, and compare the obtained vector with the learning factor Multiply them together to get the first product; in specific implementation, is equal to 0.5; the weight vector is equal to 1- The second product is obtained by multiplying; the first product is added to the second product, and the result of the addition is used as the new weight vector, which is calculated as follows:

[0126] (12)

[0127] in, is the weight vector before updating, is the weight vector after update;

[0128] S38, determine whether n is greater than the total number of signals to be sorted N, if not greater than N, add 1 to n, and then jump to S33; otherwise, the iterative loop ends and jumps to S39;

[0129] S39, sending the updated category matrix C outputted from S36 to the category merging module, performing the category merging operation, and outputting the final sample classification result;

[0130] The category merging operation is as follows:

[0131] S391, counting the number of input samples contained in each category in the category matrix C, and recording the number of categories in the C matrix as P;

[0132] S392, set the iteration loop variable p, and set the initial loop variable to p=1;

[0133] S393, determine whether the number of samples in the pth category exceeds the threshold. If so, jump to S394; if not, take the mean of the input sample set corresponding to all categories in the F2 layer as the center of each category, merge the category number and category center, and obtain a new category matrix C1, which is calculated as follows:

[0134] (14)

[0135] in, is the total number of samples, and To be merged Class and The sample set contained in the class, For the The total number of samples contained in the class, For the The total number of samples contained in the class.

[0136] S394, determine whether p is greater than the number of categories P. If not, add 1 to p and jump to S393; if greater than P, use the number of columns of the new category matrix C1 as the number of categories after merging P1, and the iterative loop ends;

[0137] S395, set the iteration loop variable p1, and set the initial loop variable to p1=1;

[0138] S396. When p1 is equal to 1, the mean of the input sample set corresponding to all categories in the category matrix C1 is taken as the center of each category, and the distance between the category centers is calculated; when p1 is greater than 1, the mean of the input sample set corresponding to all categories in the category matrix C2 is taken as the center of each category, and the distance between the category centers is calculated;

[0139] S397, determine whether the distance between the two category centers exceeds the threshold, if so, jump to S398; if not, merge the corresponding two category numbers and category centers to obtain a new category matrix C2; wherein the threshold is equal to the sum of the distances between all N sample points divided by N to the power of a and N-1 to the power of a; in specific implementation, a is equal to 0.3; the calculation is as follows:

[0140] (15)

[0141] (16)

[0142] S398, determine whether p1 is greater than the number of categories P1. If not, update p1: add 1 to p1 and jump to S396; if greater than P1, output the new category matrix C2 and the category center corresponding to each category as the final sample classification result, and the iterative loop ends;

[0143] S4, evaluating the signal classification results obtained in S3 based on the evaluation index;

[0144] The data set to be sorted is , the categories obtained after executing the algorithm are divided into , the corresponding cluster center is , the cluster division given by the reference model is , the corresponding cluster center is ;

[0145] The evaluation indicators include the number of clusters, cluster quality, centroid index and running time.

[0146] S41. Calculate cluster quality CQ, as follows

[0147] (17)

[0148] in, To classify Belongs to the original category The larger the value, the more samples can be correctly classified and the higher the accuracy. When only samples from the same category are included in .

[0149] S42. Calculate the centroid index CI

[0150] The centroid index CI is a novel indicator for evaluating whether clustering is correct or not. This indicator characterizes the matching degree of clustering before and after division based on the centroid of the category. The specific calculation formula is as follows:

[0151] (18)

[0152] in, It means that the two categories have the same structure and are completely matched. The larger the CI value, the more the sample does not correspond to the true category after classification.

[0153] So far, from S1 to S4, the radar signal sorting method based on the improved fuzzy adaptive resonance network has been completed.

[0154] In order to more effectively illustrate the above analysis, simulation experiments were conducted on the sorting results of the Fuzzy ART network without using the Visual Assessment of Cluster Tendency (VAT) algorithm (the order of the input sequence sent to the network is generated according to TOA or randomly generated) and the sorting results of the Fuzzy ART network after using the VAT algorithm. The parameters involved in the network are: selection factor =0.1, learning factor =0.5, warning threshold =0.756. Among them, , The value is a constant value, and the value of the warning threshold is determined according to the best sorting result obtained by traversing the network in the range of [0,1] with a step size of 0.002. The specific results are shown in Table 2 below:

[0155] Table 2 The impact of different input orders on network sorting results

[0156]

[0157] As can be seen from Table 2 above, the network is sensitive to the input order of samples, and under the same parameter settings, the network obtains fewer clusters for the corrected input order, which is closer to the actual situation. Although the sorting accuracy is higher after the input order generated by TOA is sent to the network, this is due to the refinement of categories, and its essence is an inaccurate classification result. According to the above simulation results, in subsequent experiments, the VAT algorithm will be used as a preprocessor, and then the corrected input will be sent to the network for sorting.

[0158] In order to illustrate the impact of the warning threshold on the sorting performance of the Fuzzy ART network, the number and quality of clusters obtained by network sorting under different warning thresholds, that is, the sorting accuracy, are simulated. Among them, the warning threshold varies from 0.65 to 0.85, with a step size of 0.002. The specific results are shown in Figure 4 .

[0159] In a complex electromagnetic environment, it is difficult to obtain the optimal warning threshold in real time, and the pulse signals intercepted by the receiver are dynamically changing, which increases the difficulty of determining In order to improve the shortcomings of the standard Fuzzy ART network, the present invention adopts an adaptive warning threshold, which is not only regulated by the current input signal, but also controlled by the previous clustering results, and comprehensively considers the distribution of category features. At the same time, considering the impact of sparse points on network training, a class merging operation is added. The above operation avoids the problem of class increment to a certain extent, and the network does not need to be manually set in advance. The number of clusters can be correctly determined when the value is large, and the sorting accuracy is high.

[0160] Figure 5 The three-dimensional characteristic distribution of radar signals after VAMFA network sorting is shown. Among them, the points represent pulse signals with sorting errors. In order to more clearly show the sorting performance of the VAMFA network on different radiation sources, the specific sorting results are listed in Table 3:

[0161] Table 3 Sorting results of VAMFA network

[0162]

[0163] As shown in Table 3 above, the misclassified signals are mainly concentrated in Groups 9-10 and 13-15. According to the corresponding three-dimensional characteristic parameters of the signals in Table 1, the dynamic ranges of the characteristic parameters of these groups of radiation source signals overlap seriously. Although the network can generally achieve correct classification, it is difficult to process the edge signals of each category in the seriously overlapping area. However, compared with the original Fuzzy ART network, which has a sorting accuracy of only about 80% when achieving the correct classification number, the performance of the VAMFA network has been greatly improved, by about 12%. This also verifies that the VAMFA algorithm can better handle complex radar signal feature sets.

[0164] In order to illustrate the effectiveness of the proposed method, several newer Fuzzy ART neural network improved algorithms are simulated and compared, including: VAT_FA, IFART, CVIFA, DDVFA_M and V_DDVFA. In order to ensure the comparison with the optimal effect of each algorithm, the specific parameter settings of each comparison algorithm are shown in Table 4. Among them, and Represents the minimum and maximum values ​​of the warning threshold. The value of the warning threshold is determined based on the best sorting result obtained by traversing the algorithm in the range [0,1] with a step size of 0.002.

[0165] Table 4 Parameter settings of different algorithms

[0166]

[0167] The above-mentioned comparison algorithms mainly improve the Fuzzy ART neural network from three aspects: warning threshold, weight vector update method and network structure. In this experiment, the number of clusters, running time, cluster quality CQ index and centroid index CI index are used to evaluate the sorting effect. The CQ index represents the sorting accuracy. Through simulation experiments, the sorting results are shown in Table 5 below:

[0168] Table 5 Comparison of sorting effects of different algorithms on radar feature sets

[0169]

[0170] Judging from the above simulation results, compared with several improved algorithms, the proposed new network model has obvious advantages: from the perspective of the number of clusters, CVIFA, DDVFA_M and V_DDVFA all have class appreciation phenomenon, and the number of categories output by the network is greater than the actual number of categories of the radiation source, and the performance of these three algorithms in other indicators is not ideal, and the sorting effect is poor; VAT_FA, IFART and VAMFA algorithms can all correctly determine the number of clusters, but the proposed VAMFA method has the highest clustering quality, reaching more than 90%; in addition, from the CI indicator, it can be seen that the classification results of this method are completely consistent with the actual category distribution structure, but VAT_FA and IFART algorithms have the phenomenon of multiple or wrong classification of one category, which is also the reason for their low sorting accuracy; finally, the running time of the VAMFA network is also relatively shorter.

[0171] In summary, when processing unknown pulse sequences with overlapping characteristic parameters, the VAMFA network can not only ensure the correct number of classifications, but also achieve sufficiently high clustering quality; it can not only improve the shortcomings affected by the input order, but also adaptively adjust the warning threshold; it can not only avoid the influence of sparse sample points, but also improve the class value-added problem, and can better adapt to complex and unknown electromagnetic environments.

[0172] Figure 1 This is a three-dimensional feature distribution diagram of radar signals of the radar signal sorting method based on the improved fuzzy adaptive resonance network of the present invention. It can be seen intuitively from the figure that the dynamic ranges of multiple signal feature parameters overlap seriously and are closely distributed, which is also a major reason why traditional clustering sorting methods are difficult to handle. Such a signal environment is also more conducive to verifying the ability of the VAMFA algorithm to solve this dilemma.

[0173] Figure 4 The schematic diagram of the network sorting performance of the radar signal sorting method based on the improved fuzzy adaptive resonance network under different warning thresholds of the present invention; it can be seen from the figure that the warning threshold The size of has a great influence on the sorting performance of the network. As the value increases, the number of clusters output by the network gradually increases. When it increases to a certain extent, it will lead to the refinement of sample classification and the sorting accuracy will also be improved to a certain extent. As can be seen from the figure, when the number of clusters is equal to the total number of input signal categories 15, The value varies between 0.738 and 0.756, and the accuracy varies between 72% and 84%, indicating that the sorting accuracy is The value is more sensitive, even if Even a small change in the value will cause a large fluctuation in the sorting performance. In addition, the average sorting accuracy is around 80%, indicating that the network is affected by the uneven distribution of sample morphology and some sparse feature points, resulting in misclassification. Therefore, in the process of network training, a suitable Value is very important.

[0174] Figure 5 This is the signal sorting result of the radar signal sorting method based on the improved fuzzy adaptive resonance network of the present invention. It can be seen from the figure that there are only a very small number of erroneous points, and they are mainly concentrated in some signals with high similarity of characteristic parameters, that is, category edge samples located in the overlapping area, which also shows that the VAMFA network can achieve accurate classification for most pulse signals.

[0175] The above is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the drawings. Any equivalent or modification completed without departing from the spirit disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A radar signal sorting method based on an improved fuzzy adaptive resonance network, characterized in that: The following steps are involved: S1, constructing a radar mixed pulse sequence, building a feature set based on characteristic parameters and normalizing it to obtain a normalized feature set; The normalized feature set includes N samples; S2, construct an improved fuzzy adaptive resonance network; The improved fuzzy adaptive resonance network includes a preprocessing module, an F0 layer, an F1 layer, an F2 layer and a class merging module; wherein the F0 layer is an input layer, the F1 layer is a membership calculation layer, and the F2 layer is a class representation layer; Among them, the F1 layer and the F2 layer are connected by a weight vector W with a dimension of 2D; S3, input all samples in the normalized feature set into the improved fuzzy adaptive resonance network constructed in S2 for processing, and output the sample classification results, which specifically includes the following sub-steps: S31, the preprocessing module preprocesses all N samples in the normalized feature set, reorders them according to the similarity between the samples, performs a visual evaluation of the clustering trend, and obtains the reordered input samples; S32, initializing the iteration loop variable n to 1, the category matrix C to be empty, and all elements of the weight vector to be 1; S32: the category matrix C is the output of the improved fuzzy adaptive resonance network, and the rows in C represent the numbers of the input samples contained in different categories, and the columns represent the number of categories; S33, sending the nth sample of the input samples reordered in S31 to the F0 layer to perform a complement coding operation to obtain a sample complement, and concatenating the sample complement with the nth sample to obtain an extended input sample; S34, sending the input sample expanded by S33 and the weight vector from the top to the bottom of the F2 layer to the F1 layer, performing a mode selection operation, calculating the membership of the input sample and the neuron, and selecting the F2 layer neuron corresponding to the maximum membership as the winning neuron; The weight vector corresponding to the winning neuron is recorded as W J ; S35, the weight vector W corresponding to the input sample expanded by S33 and the winning neuron selected by S34 J Send it to the F2 layer to perform pattern matching operation, that is, calculate the matching degree between the input sample and the winning neuron; S36: Determine whether the matching degree calculated in S35 exceeds a warning threshold, where the warning threshold is , if it is greater than or equal to the warning threshold, the reset node sends a resonance signal, the reset node is recorded as R, the network starts learning, the nth input sample is added to the input samples contained in the corresponding category in the C matrix, and the weight vector is updated, and the process jumps to S37, otherwise if it is less than the warning threshold, the network mismatch is reset, the reset node sends a reset signal, and the network re-searches for neurons that meet the matching degree exceeding the warning threshold; if the matching degree of all neurons in the F2 layer and the input samples cannot exceed the warning threshold, the nth input sample is added to the F2 layer as a new neuron, and the nth input sample is used as a new weight vector and added to the category matrix C to obtain the updated category matrix C; The calculation of the warning threshold is specifically as follows: S361, calculate the fuzzy distance between the weight vector WJ corresponding to the winning node and the input sample in the kth dimension as the reference distance, denoted as , where k ranges from [1,2D]; the fuzzy distance is equal to 1 minus the inverse exponential of the absolute value of the weight vector WJ minus the value of the input sample on the kth dimension; where the fuzzy factor is denoted by m and ranges from 0 to 20; S362, calculate the fuzzy distance between the weight vector WJ corresponding to the winning node and all classified input samples belonging to the winning node in the kth dimension as the weighted distance, denoted as , where the weighted distance is equal to the fuzzy distance between each input sample and the weight vector WJ multiplied by the value of the input sample in the kth dimension, divided by the sum of the elements of all input samples in the kth dimension; S363, comparing the reference distance obtained in S361 and the weighted distance obtained in S362, and taking the smaller value between the two as the value of the warning threshold in the kth dimension; S364: Select the maximum value of all elements of the warning threshold as the final warning threshold for matching judgment ; In S37, the learning factor is recorded as , the value range is [0,1]; S37, update the weight vector, specifically take the minimum value of the input sample and the element in each dimension of the weight vector, and compare the obtained vector with the learning factor Multiply to get the first product; weight vector and 1- The second product is obtained by multiplying; the first product is added to the second product, and the result of the addition is used as a new weight vector; S38, determine whether n is greater than the total number of signals to be sorted N, if not greater than N, add 1 to n, and then jump to S33; otherwise, the iterative loop ends and jumps to S39; S39, sending the updated category matrix C outputted from S36 to the category merging module, performing the category merging operation, and outputting the final sample classification result; S4, evaluating the sample classification results obtained in S3 based on the evaluation indicators; Among them, the evaluation indicators include the number of clusters, cluster quality, centroid index and running time.

2. The radar signal sorting method according to claim 1, characterized in that: S1 specifically includes the following sub-steps: S11, construct radar mixed pulse sequence; S12, collecting signal pulses to be sorted from the radar mixed pulse sequence constructed in S11, and recording the total number of signal pulses to be sorted as N; S13, forming a feature set from the feature parameters of each signal pulse to be sorted, and performing normalization processing on the feature set, that is, changing each dimensional feature in the feature set to the same scale range to obtain a normalized feature set.

3. The radar signal sorting method according to claim 2, characterized in that: The normalization processing described in S1 adopts maximum and minimum normalization; the types of radar mixed pulse sequences described in S11 are M; each radar includes several signal pulses and the dimension of each signal pulse is D, that is, the signal pulse is represented by D characteristic parameters; in S13, the total number of feature sets is the same as the total number of signal pulses to be sorted, both are N; the characteristic parameters of each signal pulse in the feature set are regarded as N samples.

4. The radar signal sorting method according to claim 1, characterized in that: The preprocessing module S31 preprocesses all N samples in the normalized feature set, reorders them according to the similarity between the samples, and performs a visual evaluation of the clustering trend to obtain the reordered input samples, specifically: Initialize the index matrix and set it to empty; calculate the distance dissimilarity matrix between samples, take one of the two samples with the farthest distance as the starting point, and add its corresponding position in the feature set to the index matrix; according to the minimum distance to any sample in the current index matrix, iterate according to the single-link hierarchical clustering and add the corresponding positions of the remaining samples in the feature set until all samples are traversed to obtain the final index matrix; rearrange the order of samples according to the index matrix to obtain the reordered input samples.

5. The radar signal sorting method according to claim 1, characterized in that: The dimension of the expanded input sample in S33 is 2D, that is, the output of the F0 layer; the complement coding operation is specifically: finding the complement code for each dimension of the sample to obtain the sample complement code.

6. The radar signal sorting method according to claim 1, characterized in that: The mode selection operation in S34 is specifically as follows: taking the minimum value of the elements in each dimension of the input sample expanded in S33 and the weight vector of each connected neuron, and calculating the 1 norm of the obtained vector as the numerator; Calculate the 1 norm of each weight vector and add it to the selection factor as the denominator; calculate the ratio of the numerator to the denominator to get the membership degree; then select the F2 layer neuron corresponding to the maximum membership degree as the winning neuron; where the selection factor is recorded as , the value range is 0 to 1.

7. The radar signal sorting method according to claim 1, characterized in that: The pattern matching operation in S35 is specifically: taking the nth input sample and the weight vector W J The minimum value of the elements in each dimension, calculate the 1 norm of the obtained vector as the numerator, and the 1 norm of the input sample as the denominator; calculate the ratio of the numerator to the denominator, that is, the matching degree of the input sample and the winning neuron.

8. The radar signal sorting method according to claim 1, characterized in that: The category merging operation described in S39 is specifically as follows: S391, counting the number of input samples contained in each category in the category matrix C, and recording the number of categories in the C matrix as P; S392, set the iteration loop variable p, and set the initial loop variable to p=1; S393, determine whether the number of samples in the pth category exceeds the threshold. If not, take the mean of the input sample set corresponding to all categories in the F2 layer as the center of each category, merge the category number and the category center, and obtain a new category matrix C1; if it exceeds, jump to S394; wherein the threshold is equal to N divided by P divided by 2; S394, determine whether p is greater than the number of categories P. If not, add 1 to p and jump to S393; otherwise, use the number of columns of the new category matrix C1 as the number of categories after merging P1, and the iterative loop ends; S395, set the iteration loop variable p1, and set the initial loop variable to p1=1; S396. When p1 is equal to 1, the mean of the input sample set corresponding to all categories in the category matrix C1 is taken as the center of each category, and the distance between the category centers is calculated; when p1 is greater than 1, the mean of the input sample set corresponding to all categories in the category matrix C2 is taken as the center of each category, and the distance between the category centers is calculated; S397, judging whether the distance between the two category centers exceeds a threshold value, if not, merging the two corresponding category numbers and category centers to obtain a new category matrix C2; if more, jumping to S398; S398, determine whether p1 is greater than the number of categories P1. If not, add 1 to p1 and jump to S396; if greater than P1, output the new category matrix C2 and the category center corresponding to each category as the final sample classification result, and the iterative loop ends.

9. The radar signal sorting method according to claim 8, characterized in that: In S397, the threshold is equal to the sum of the distances between all N sample points divided by N to the power of a and N-1 to the power of a, and the value range of a is 0 to 1.

Images