A method for signal sorting and identification in complex electromagnetic environments

Abstract

This invention relates to a signal sorting and identification method in complex electromagnetic environments, comprising: preprocessing the electromagnetic pulse attribute features in a real-time read-in electromagnetic pulse descriptor sequence, including denoising and normalization; performing preliminary sorting of the preprocessed electromagnetic pulse descriptor sequence based on DBSCAN density clustering to obtain a first electromagnetic signal cluster; performing preliminary merging of the first electromagnetic signal cluster based on pulse repetition frequency matching to obtain a second electromagnetic signal cluster; further merging the second electromagnetic signal cluster based on pulse timing continuity judgment to obtain a third electromagnetic signal cluster; performing cluster association on the third electromagnetic signal cluster based on average Euclidean distance between clusters to obtain a fourth electromagnetic signal cluster; and performing target identification on the fourth electromagnetic signal cluster based on a random forest model to obtain the target type to which each cluster in the fourth electromagnetic signal cluster belongs. This invention improves the accuracy of signal sorting, tracking, and target identification.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic signal sorting and recognition technology, and to a method for signal sorting and recognition in complex electromagnetic environments. Background Technology

[0002] With the continuous development and advancement of modern technology, various information devices are constantly entering fields such as transportation and communication. Numerous, complex, densely overlapping, and dynamically interacting electromagnetic signals permeate the electromagnetic environment, making it increasingly complex. Accurate target positioning and tracking algorithms are a crucial pillar in the current field of traffic navigation. How to accurately acquire as much electromagnetic information about the target as possible is key to achieving positioning and tracking tasks. Electromagnetic signal sorting and recognition, as an important component of electromagnetic tracking technology, is the prerequisite and foundation for feature extraction from electromagnetic targets. This technology mainly consists of two parts: electromagnetic signal sorting and electromagnetic signal recognition.

[0003] As a core technology, electromagnetic signal sorting aims to classify received electromagnetic pulses based on their different electromagnetic characteristics. In today's complex electromagnetic environment, with severe signal overlap and frequent radar mode switching, traditional single-feature sorting algorithms based on pulse interval time (PRI) struggle to accurately classify electromagnetic pulses from different individuals, easily leading to misclassification and omissions. Multi-feature sorting algorithms based on density clustering comprehensively consider the multi-dimensional features of electromagnetic pulses, enabling accurate sorting based on the density relationship between them. This allows for the classification of electromagnetic pulses from different individuals and operating modes into different clusters with low latency. Subsequently, a merging algorithm based on end-to-end connection and PRI matching is used to merge clusters of the same individual but different operating modes. Simultaneously, to ensure effective tracking of individual targets, a cluster association algorithm based on average Euclidean distance can be used to achieve stable tracking. Electromagnetic signal recognition primarily addresses the identification of electromagnetic equipment models. Leveraging existing datasets and prior knowledge, a random forest based on multi-dimensional features is trained, and accurate signal recognition can be achieved through multi-decision tree ensemble prediction.

[0004] Accurate and reliable sorting and identification results are the prerequisite and guarantee for in-depth mining of electromagnetic information. Full utilization of multi-dimensional features can achieve complementary advantages between information, greatly improving the accuracy and rationality of the results. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to disclose a signal sorting and recognition method in complex electromagnetic environments, thereby solving the accuracy problem of signal sorting and tracking and target recognition.

[0006] This invention discloses a signal sorting and identification method for complex electromagnetic environments, comprising:

[0007] Step S1: Perform preprocessing on the electromagnetic pulse attribute features in the real-time read electromagnetic pulse description word sequence, including denoising and normalization.

[0008] Step S2: Based on DBSCAN density clustering, perform preliminary sorting on the preprocessed electromagnetic pulse descriptor sequence to obtain the first electromagnetic signal cluster;

[0009] Step S3: Based on pulse repetition frequency matching, the first electromagnetic signal cluster is initially merged to obtain the second electromagnetic signal cluster;

[0010] Step S4: Based on the pulse timing continuity judgment, the second electromagnetic signal cluster is re-merged to obtain the third electromagnetic signal cluster;

[0011] Step S5: Based on the average Euclidean distance between clusters, perform cluster correlation on the third electromagnetic signal cluster to obtain the fourth electromagnetic signal cluster;

[0012] Step S6: Based on the random forest model, target identification is performed on the fourth electromagnetic signal cluster to obtain the target type of each cluster in the fourth electromagnetic signal cluster.

[0013] Furthermore, the density clustering process based on DBSCAN in step S2 includes:

[0014] Step S2-1: For the preprocessed data, the electromagnetic pulse is mapped into a point in three-dimensional space using the normalized carrier frequency, pulse width, and pulse amplitude as parameter features.

[0015] Step S2-2: In three-dimensional space, use a recursive method to find cluster centers that meet the requirements of neighborhood radius and minimum number of points in a single cluster; determine a cluster for each cluster center and establish a corresponding cluster information database;

[0016] Step S2-3: Add points to the cluster according to the density reachability requirement, and store the pulse information corresponding to the points in the corresponding cluster information database until no density reachable points can be found to end the loop and generate a cluster.

[0017] After clustering each cluster center, the first electromagnetic signal cluster is obtained according to different individuals and different radar operating modes.

[0018] Furthermore, the preliminary merging process of the first electromagnetic signal cluster in step S3 includes:

[0019] Step S3-1: For each pulse in the first electromagnetic signal cluster, arrange the pulses in ascending order according to the magnitude of the pulse arrival time, and calculate the pulse repetition frequency of each pulse.

[0020] Step S3-2: Based on the histogram of pulse repetition frequencies of each cluster, calculate the representative pulse repetition frequency of each cluster;

[0021] Step S3-3: Compare the representative pulse repetition frequencies of each cluster in the first electromagnetic signal cluster pairwise to calculate the similarity. Merge the clusters that meet the similarity merging requirements to obtain the second electromagnetic signal cluster.

[0022] Furthermore, the calculation process for the representative pulse repetition frequency of each cluster in step S3-2 includes:

[0023] 1) Construct a histogram of pulse repetition frequencies; the horizontal axis of the histogram represents the pulse repetition frequency value, and the vertical axis represents the frequency count; the histogram shows the pulse repetition frequency values ​​arranged in ascending order, and the frequency of occurrence of each pulse repetition frequency value in the cluster is counted.

[0024] 2) Set a sliding window with a fixed window width and sliding step size in the histogram; use the sliding window to traverse the histogram starting from the leftmost side and following the sliding step size; calculate the sum of frequencies within the sliding window before each slide; and store the position of the sliding window and the corresponding sum of frequencies in the frequency comparison library.

[0025] 3) After the traversal is completed, find the sliding window index and corresponding sliding window position of the maximum frequency of each cluster in the frequency comparison library, and use the frequency value in the sliding window at that position to perform a weighted sum of the pulse repetition frequency in the window as the representative pulse repetition frequency of each cluster.

[0026] Furthermore, the formula for calculating the pulse repetition frequency is as follows:

[0027]

[0028] Let pri be the representative pulse repetition frequency of the nth cluster in the i-th time slice; i,n,k This is a weighted summation of the k-th slide of the sliding window in the n-th cluster within the i-th time slice; pri i,n,k,l Let pri,num be the frequency of the l-th pulse covered by the sliding window of the n-th cluster in the i-th time slice during the k-th sliding motion. i,n,k,l num represents the frequency corresponding to the time difference. i,n,k This represents the total frequency covered by the k-th sliding window range of the n-th cluster in the i-th time slice.

[0029] Furthermore, in step S3-3, the formula for calculating similarity is:

[0030]

[0031] sim a,b This represents the similarity between cluster a and cluster b. Let be the pulse repetition frequency of the a-th cluster. Let be the pulse repetition frequency of the b-th cluster;

[0032] When the similarity value is less than the set threshold pri_threshold, all pulses in cluster a are merged into cluster b.

[0033] Furthermore, the process of re-merging the second electromagnetic signal cluster in step S4 includes:

[0034] Step S4-1: Arrange the pulses of each cluster in the second electromagnetic signal cluster in ascending order of pulse arrival time, determine the continuity of pulses in each cluster, and after finding a cluster with a breakpoint, split the cluster from the breakpoint to form two clusters, thus obtaining an electromagnetic signal cluster with continuous time within the cluster.

[0035] Step S4-2: Calculate the pulse repetition frequency and representative pulse repetition frequency of the cluster head pulse and the cluster tail pulse of each cluster in a time-continuous electromagnetic signal cluster.

[0036] Step S4-3: Calculate the arrival time difference between the tail pulse of each cluster and the head pulse of all other clusters; based on the time difference, merge the two clusters that meet the merging tolerance requirements into one cluster in sequence; after the cluster merging, the third electromagnetic signal cluster is obtained.

[0037] Furthermore, the tolerance condition for the difference is:

[0038] diff a,b ≥0

[0039]

[0040] Among them, diff a,b The arrival time difference is the pulse tail of cluster a and the pulse head of cluster b. Let be the representative pulse repetition frequency of the cluster tail pulse of the a-th cluster; is the representative pulse repetition frequency of the cluster head pulse of the b-th cluster; tolerance is the difference tolerance; time_a and time_b are arbitrary integers from 0 to 5, and not both 0 at the same time; when any combination of the values ​​of time_a and time_b satisfies the difference tolerance condition, the cluster tail of the a-th cluster and the cluster head of the b-th cluster are connected and merged into one cluster.

[0041] Furthermore, the cluster association process for the third electromagnetic signal cluster in step S5 includes:

[0042] Step S5-1: Calculate the average Euclidean distance between two clusters in the third electromagnetic signal cluster;

[0043] The formula for calculating the average Euclidean distance is:

[0044]

[0045] Among them, dist a,b Cluster B for the i-th time slice i,a Cluster B of the (i+1)th time slice i+1,b ;p i,a,j For cluster B i,a Point J is a given point in p. i,a,j In cluster B i,a The position of p in the middle i+1,b,k For cluster B i+1,b Points in the middle, k is p i+1,b,k In cluster B i+1,b The position in the middle;

[0046] Step S5-2: Determine whether the average Euclidean distance between clusters in the third electromagnetic signal cluster is less than the maximum associatable average Euclidean distance tolerance; if yes, then associate these two clusters as electromagnetic signal clusters of the same target individual.

[0047] Furthermore, when training the random forest model used in target recognition, the CART decision tree algorithm is used to construct decision trees. When constructing each decision tree, a subset is randomly sampled from the training data for training to ensure the diversity between decision trees.

[0048] When using the random forest model for prediction, the electromagnetic pulse descriptor sequence read in real time is processed through steps S1-S5 to obtain the fourth electromagnetic signal cluster. The signal data of the fourth electromagnetic signal cluster is then input into the random forest model. Each decision tree in the random forest model completes the predefined classification task based on the input attribute features and outputs the classification result of each cluster. Finally, the target type of each cluster is obtained through voting ensemble, thus achieving the target recognition task.

[0049] This invention can achieve one of the following beneficial effects:

[0050] 1. This invention mines the multidimensional features of electromagnetic signals, combines the real-time read electromagnetic pulse information into the input features of the model, and performs preprocessing steps such as outlier removal and normalization to avoid affecting the correctness of subsequent sorting and identification, thereby improving the accuracy of the results.

[0051] 2. This invention employs a signal sorting algorithm that first breaks down the clusters and then merges them. First, a density clustering method based on DBSCAN is used for initial sorting, generating numerous small clusters with extremely high purity. Then, feature matching is performed using carrier frequency, pulse width, and pulse amplitude characteristics to complete initial cluster merging, which significantly reduces batch expansion while maintaining high merging accuracy. Finally, time continuity is utilized to perform cluster merging based on a head-to-tail connection algorithm, achieving the goal of sorting according to the target individuals.

[0052] 3. This invention employs a cross-time-slice association algorithm based on the average Euclidean distance between clusters, which effectively improves the tracking performance of individual targets and provides stable support for subsequent electromagnetic feature analysis algorithms.

[0053] 4. This invention utilizes the random forest algorithm to achieve the target recognition task, and by leveraging the iterative update capability of machine learning, it greatly improves the accuracy of target type recognition. Attached Figure Description

[0054] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0055] Figure 1 This is a flowchart of a signal sorting and identification method under complex electromagnetic environment in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the representative pulse repetition frequency calculated based on histogram in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of the average Euclidean distance between clusters in an embodiment of the present invention. Detailed Implementation

[0058] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0059] One embodiment of the present invention discloses a signal sorting and identification method in a complex electromagnetic environment, such as... Figure 1 As shown, it includes:

[0060] Step S1: Perform preprocessing on the electromagnetic pulse attribute features in the real-time read electromagnetic pulse description word sequence, including denoising and normalization.

[0061] Step S2: Based on DBSCAN density clustering, perform preliminary sorting on the preprocessed electromagnetic pulse descriptor sequence to obtain the first electromagnetic signal cluster;

[0062] Step S3: Based on pulse repetition frequency matching, the first electromagnetic signal cluster is initially merged to obtain the second electromagnetic signal cluster;

[0063] Step S4: Based on the pulse timing continuity judgment, the second electromagnetic signal cluster is re-merged to obtain the third electromagnetic signal cluster;

[0064] Step S5: Based on the average Euclidean distance between clusters, perform cluster correlation on the third electromagnetic signal cluster to obtain the fourth electromagnetic signal cluster;

[0065] Step S6: Based on the random forest model, target identification is performed on the fourth electromagnetic signal cluster to obtain the target type of each cluster in the fourth electromagnetic signal cluster.

[0066] Specifically, the preprocessing in step S1 includes:

[0067] Step S1-1: Read the electromagnetic pulse descriptor sequence sequentially in real time, and store the attribute features of the read electromagnetic pulse descriptor into the corresponding feature library according to the sequence number;

[0068] The attribute features include pulse attribute features such as pulse arrival time (TOA), carrier frequency (RF), pulse width (PW), and pulse amplitude (PA).

[0069] More specifically, the pulse descriptor sequence T i It is a sequence of pulse descriptors arranged according to the pulse arrival time. In most cases, the signal receiver processes the received signal periodically to generate pulse descriptors for the corresponding time slice. Here, i represents the index number of different time slices in the pulse descriptor sequence. To ensure the timeliness of the algorithm, except for cluster association algorithms based on average Euclidean distance, subsequent processing is performed on the pulse descriptor information within a certain time slice.

[0070] The pulse descriptor sequence can be represented as T i = <p i,1 ,…,p i,k ,…,p i,m >;

[0071] Where, p i,k =(toa i,k ,rf i,k ,pw i,k ,pa i,k ) indicates time. i,k The arrival time of the received pulse toa i,k Carrier frequency rf i,k Pulse width pw i,k , pulse amplitude pa i,k ;

[0072] time i,k <time i,k+1 ; and m represents the length of the pulse sequence in this time slice; the start time of the sequence is time. i,1 The end time is time i,m .

[0073] Step S1-2: Remove outlier data from attribute features;

[0074] Due to signal receiver errors and environmental noise, the attribute values ​​of a small number of pulse descriptors deviate significantly from the other attribute values ​​of the same label.

[0075] Therefore, in this embodiment, outliers exceeding the threshold are identified and removed by using three times the standard deviation of the attribute values ​​with the same label.

[0076] Step S1-3: Normalize the attribute features using discrete standardization to obtain the preprocessed electromagnetic pulse attribute features;

[0077] The normalization formula is:

[0078]

[0079] Where x represents the original attribute feature value of the current pulse descriptor, min represents the minimum value of the corresponding label attribute in the feature library, max represents the maximum value of the corresponding label attribute in the feature library, and x′ represents the attribute feature value after normalization, with a range of [0,1].

[0080] Specifically, the density clustering process based on DBSCAN in step S2 includes:

[0081] Step S2-1: For the preprocessed data, the electromagnetic pulse is mapped into a point in three-dimensional space using the normalized carrier frequency, pulse width, and pulse amplitude as parameter features.

[0082] Step S2-2: In three-dimensional space, use a recursive method to find cluster centers that satisfy the requirements of neighborhood radius eps and minimum number of points threshold_num for a single cluster; determine a cluster for each cluster center and establish a corresponding cluster information database;

[0083] Step S2-3: Add points to the cluster according to the density reachability requirement, and store the pulse information corresponding to the points in the corresponding cluster information database until no density reachable points can be found to end the loop and generate a cluster.

[0084] After clustering each cluster center, the first electromagnetic signal cluster is obtained according to different individuals and different radar operating modes.

[0085] In the first electromagnetic signal cluster, the cluster center is c. i,n A cluster B i,n The pulse information p corresponding to the midpoint i,n,j =(toa i,n,j ,rf i,n,j ,pw i,n,j ,pa i,n,j );

[0086] Among them, toa i,n,j ,rfi,n,j ,pw i,n,j pa i,n,j These represent the arrival time, carrier frequency, pulse width, and pulse amplitude of the j-th pulse in the n-th cluster within the i-th time slice, respectively.

[0087] In the cluster partitioning in step S2-2, the smaller the neighborhood radius is set, the higher the purity of each cluster, but at the same time, the batching phenomenon will be more serious. Therefore, when the neighborhood radius is set to be small, it is necessary to merge the batched clusters.

[0088] Specifically, in step S3, the initial cluster merging process of the first electromagnetic signal cluster includes:

[0089] Step S3-1: For each pulse in the first electromagnetic signal cluster, arrange the pulses in ascending order according to the magnitude of the pulse arrival time, and calculate the pulse repetition frequency pri of each pulse.

[0090] After arranging the pulses in ascending order, the arrival time difference between adjacent pulses is calculated; each arrival time difference is used as the pulse repetition frequency of the later pulse in the adjacent pulses.

[0091] The formula is: pri i,n,j =toa i,n,j -toa i,n,j-1 ;pri i,n,j It represents the time difference between the (j-1)th pulse and the jth pulse in the nth cluster of the i-th time slice.

[0092] Step S3-2: Based on the histogram of the pulse repetition frequency pri for each cluster, calculate the representative pulse repetition frequency for each cluster.

[0093] Specifically, including:

[0094] 1) Construct a histogram of pulse repetition frequencies; the horizontal axis of the histogram represents the pulse repetition frequency value, and the vertical axis represents the frequency count; the histogram shows the pulse repetition frequency values ​​arranged in ascending order, and the frequency of occurrence of each pulse repetition frequency value in the cluster is counted.

[0095] Preferably, the frequency of occurrence of the pulse repetition frequency value can be calculated using Counter.

[0096] 2) Set a sliding window with a fixed window width and sliding step size in the histogram; use the sliding window to traverse the histogram from the leftmost side according to the sliding step size; calculate the sum of frequencies within the sliding window before each slide; and store the position of the sliding window and the corresponding sum of frequencies in the frequency comparison library.

[0097] 3) After the traversal is completed, find the sliding window index and corresponding sliding window position of the maximum frequency of each cluster in the frequency comparison library, and use the frequency value in the sliding window at that position to perform a weighted sum of the pulse repetition frequency in the window as the representative pulse repetition frequency of each cluster.

[0098] The formula representing the pulse repetition frequency is as follows:

[0099]

[0100] Let pri be the representative pulse repetition frequency of the nth cluster in the i-th time slice; i,n,k This is a weighted summation of the k-th slide of the sliding window in the n-th cluster within the i-th time slice; pr i,n,k,l Let pri,num be the frequency of the l-th pulse covered by the sliding window of the n-th cluster in the i-th time slice during the k-th sliding motion. i,n,k,l num represents the frequency corresponding to the time difference. i,n,k This represents the total frequency covered by the k-th sliding window range of the n-th cluster in the i-th time slice.

[0101] like Figure 2 As shown, a schematic diagram of the representative pulse repetition frequency calculated based on the histogram is given;

[0102] Step S3-3: Compare the representative pulse repetition frequencies of each cluster in the first electromagnetic signal cluster pairwise to calculate the similarity. Merge the clusters that meet the similarity merging requirements to obtain the second electromagnetic signal cluster.

[0103] The formula for calculating similarity is as follows:

[0104]

[0105] sim a,b This represents the similarity between cluster a and cluster b. Let be the pulse repetition frequency of the a-th cluster. Let be the pulse repetition frequency of the b-th cluster;

[0106] When the similarity value is less than the set threshold pri_threshold, all pulses in cluster a are merged into cluster b; this process is repeated recursively until all clusters that meet the merging requirements are merged to obtain the second electromagnetic signal cluster.

[0107] After the initial cluster merging in step S3, the batching phenomenon can be greatly reduced. However, there are still some related clusters that can be merged. These clusters have strong temporal continuity. Therefore, the temporal relationship between related clusters can be used for further merging.

[0108] Specifically, step S4 involves re-merging the second electromagnetic signal cluster; this includes:

[0109] Step S4-1: Arrange the pulses of each cluster in the second electromagnetic signal cluster in ascending order of pulse arrival time, determine the continuity of pulses in each cluster, and after finding a cluster with a breakpoint, split the cluster from the breakpoint to form two clusters, thus obtaining an electromagnetic signal cluster with continuous time within the cluster.

[0110] Specifically, including:

[0111] 1) Arrange the pulse clusters in ascending order of pulse arrival time, and calculate the pulse repetition frequency of each pulse in each cluster and the representative pulse repetition frequency of the cluster.

[0112] The calculation methods for the pulse repetition frequency of each pulse in the cluster and the representative pulse repetition frequency of the cluster are the same as those in steps S3-1 and S3-2. Please refer to them for details.

[0113] 2) Compare the pulse repetition frequency of each pulse in the cluster with the pulse repetition frequency of the cluster's representative pulse. When the repetition frequency of a pulse exceeds the time multiple tolerance of the pulse repetition frequency of the cluster's representative pulse, that point is considered a breakpoint, and the cluster will be broken from there, splitting into two clusters.

[0114] The time multiplier tolerance can be set according to the specific circumstances.

[0115] Step S4-2: Calculate the pulse repetition frequency and representative pulse repetition frequency of the cluster head pulse and the cluster tail pulse of each cluster in a time-continuous electromagnetic signal cluster.

[0116] The first ten pulses and the last ten pulses of each cluster are extracted as the cluster head pulse and the cluster tail pulse, respectively.

[0117] Calculate the pulse repetition frequency of each pulse in the cluster head pulse and the cluster tail pulse, as well as the representative pulse repetition frequency of the cluster head pulse and the cluster tail pulse.

[0118] The calculation method is the same as that in steps S3-1 and S3-2, please refer to the specific steps.

[0119] Step S4-3: Calculate the arrival time difference between each cluster tail pulse and all other cluster head pulses; based on the time difference, merge two clusters that meet the merging tolerance requirements into one cluster in sequence; after cluster merging, the third electromagnetic signal cluster is obtained;

[0120] The tolerance conditions for the difference are:

[0121] diff a,b ≥0

[0122]

[0123] Among them, diff a,b The arrival time difference is the pulse tail of cluster a and the pulse head of cluster b. Let be the representative pulse repetition frequency of the cluster tail pulse of the a-th cluster; is the representative pulse repetition frequency of the cluster head pulse of the b-th cluster; tolerance is the difference tolerance; time_a and time_b are arbitrary integers from 0 to 5, and not both 0 at the same time; when any combination of the values ​​of time_a and time_b satisfies the difference tolerance condition, the cluster tail of the a-th cluster and the cluster head of the b-th cluster are connected and merged into one cluster.

[0124] The above algorithm in a single time slice T i Within the given timeframe, the electromagnetic pulse sorting task has been completed for each individual target. In most cases, a single target will continuously emit electromagnetic pulses across multiple time slices. Correlating the electromagnetic signals of an individual target across time slices is crucial for achieving individual tracking. This involves linking the electromagnetic signals of different clusters (B) from different time slices... i,n The electromagnetic information features contained therein are stored in different libraries.

[0125] Specifically, the cluster association process for the third electromagnetic signal cluster in step S5 includes:

[0126] Step S5-1: Calculate the average Euclidean distance between two clusters in the third electromagnetic signal cluster;

[0127] When calculating the average Euclidean distance, all pulses of each cluster in the third electromagnetic signal cluster are normalized to the carrier frequency rf. i,k Pulse width pw i,k , pulse amplitude pa i,k The parameter features are mapped to points in three-dimensional space.

[0128] The formula for calculating the average Euclidean distance is:

[0129]

[0130] Among them, dist a,b Cluster B for the i-th time slice i,a Cluster B of the (i+1)th time slice i+1,b ;p i,a,j For cluster B i,a Point J is a given point in p. i,a,j In cluster B i,a The position of p in the middle i+1,b,k For cluster B i+1,b Points in the middle, k is p i+1,b,k In cluster B i+1,b The position in the middle;

[0131] like Figure 3As shown, a schematic diagram of the average Euclidean distance between clusters is presented.

[0132] Step S5-2: Determine whether the average Euclidean distance between clusters in the third electromagnetic signal cluster is less than the maximum associatable average Euclidean distance tolerance; if yes, then associate these two clusters as electromagnetic signal clusters of the same target individual.

[0133] The fourth electromagnetic signal cluster is obtained by associating clusters with the same target.

[0134] Specifically, in the target recognition in step S6, when training the random forest model, the CART decision tree algorithm is used to construct decision trees. When constructing each decision tree, a subset is randomly sampled from the data for training to ensure the diversity between decision trees.

[0135] The training data can be the previously obtained data from the fourth electromagnetic signal cluster, or the data processed in step S1.

[0136] When using the random forest model for prediction, the electromagnetic pulse descriptor sequence read in real time is processed through steps S1-S5 to obtain the fourth electromagnetic signal cluster. The signal data of the fourth electromagnetic signal cluster is then input into the random forest model. Each decision tree in the random forest model will complete the predefined classification task based on the input attribute features and output the classification result of each cluster. Finally, the target type of each cluster is obtained through voting ensemble, thus achieving the target recognition task.

[0137] In summary, the embodiments of the present invention mine the multidimensional features of electromagnetic signals, combine the real-time read electromagnetic pulse information into the input features of the model, and perform preprocessing steps such as outlier removal and normalization to avoid affecting the correctness of subsequent sorting and identification, thereby improving the accuracy of the results.

[0138] A signal sorting algorithm that first breaks down and then merges signals is adopted. First, a density clustering method based on DBSCAN is used for preliminary sorting to generate a large number of small clusters with extremely high purity. Then, feature matching is performed using carrier frequency, pulse width, and pulse amplitude characteristics to complete the preliminary cluster merging, which can greatly reduce the batching phenomenon while maintaining a high merging accuracy. Finally, cluster merging is performed based on the head-to-tail connection algorithm using time continuity to achieve the goal of sorting according to the target individuals.

[0139] A cross-time-slice association algorithm based on the average Euclidean distance between clusters was adopted, which effectively improved the tracking effect of individual targets and can provide stable support for subsequent electromagnetic feature analysis algorithms.

[0140] The random forest algorithm was used to implement the target recognition task. With the help of the iterative update capability of machine learning, the accuracy of target type recognition was greatly improved.

[0141] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for signal sorting and identification in complex electromagnetic environments, characterized in that, include: Step S1: Perform preprocessing on the electromagnetic pulse attribute features in the real-time read electromagnetic pulse description word sequence, including denoising and normalization. Step S2: Based on DBSCAN density clustering, perform preliminary sorting on the preprocessed electromagnetic pulse descriptor sequence to obtain the first electromagnetic signal cluster; Step S3: Based on pulse repetition frequency matching, the first electromagnetic signal cluster is initially merged to obtain the second electromagnetic signal cluster; Step S4: Based on the pulse timing continuity judgment, the second electromagnetic signal cluster is re-merged to obtain the third electromagnetic signal cluster; Step S5: Based on the average Euclidean distance between clusters, perform cluster correlation on the third electromagnetic signal cluster to obtain the fourth electromagnetic signal cluster; include: Step S5-1: Calculate the average Euclidean distance between two clusters in the third electromagnetic signal cluster; When calculating the average Euclidean distance, all pulses of each cluster in the third electromagnetic signal cluster are mapped into points in three-dimensional space using normalized carrier frequency, pulse width, and pulse amplitude as parameters. The formula for calculating the average Euclidean distance is: ； in, For the first Clusters of time slices and the Clusters of time slices The distance; For clusters The point in the middle, for In cluster The position in the middle, For clusters The point in the middle, for In cluster The position in the middle; Step S5-2: Determine whether the average Euclidean distance between clusters in the third electromagnetic signal cluster is less than the maximum associative average Euclidean distance tolerance; if yes, then associate these two clusters as electromagnetic signal clusters of the same target individual. Step S6: Based on the random forest model, target identification is performed on the fourth electromagnetic signal cluster to obtain the target type of each cluster in the fourth electromagnetic signal cluster.

2. The signal sorting and identification method under complex electromagnetic environments according to claim 1, characterized in that, The density clustering process based on DBSCAN in step S2 includes: Step S2-1: For the preprocessed data, the electromagnetic pulse is mapped into a point in three-dimensional space using the normalized carrier frequency, pulse width, and pulse amplitude as parameter features. Step S2-2: In three-dimensional space, use a recursive method to find cluster centers that meet the requirements of neighborhood radius and minimum number of points in a single cluster; determine a cluster for each cluster center and establish a corresponding cluster information database; Step S2-3: Add points to the cluster according to the density reachability requirement, and store the pulse information corresponding to the points in the corresponding cluster information database until no density reachable points can be found to end the loop and generate a cluster. After clustering each cluster center, the first electromagnetic signal cluster is obtained according to different individuals and different radar operating modes.

3. The signal sorting and identification method under complex electromagnetic environments according to claim 2, characterized in that, The preliminary merging process of the first electromagnetic signal cluster in step S3 includes: Step S3-1: For each pulse in the first electromagnetic signal cluster, arrange the pulses in ascending order according to the magnitude of the pulse arrival time, and calculate the pulse repetition frequency of each pulse. Step S3-2: Based on the histogram of pulse repetition frequencies of each cluster, calculate the representative pulse repetition frequency of each cluster; Step S3-3: Compare the representative pulse repetition frequencies of each cluster in the first electromagnetic signal cluster pairwise to calculate the similarity. Merge the clusters that meet the similarity merging requirements to obtain the second electromagnetic signal cluster.

4. The signal sorting and identification method under complex electromagnetic environments according to claim 3, characterized in that, The calculation process for the representative pulse repetition frequency of each cluster in step S3-2 includes: 1) Construct a histogram of pulse repetition frequencies; the horizontal axis of the histogram represents the pulse repetition frequency value, and the vertical axis represents the frequency count; the histogram shows the pulse repetition frequency values ​​arranged in ascending order, and the frequency of occurrence of each pulse repetition frequency value in the cluster is counted. 2) Set a sliding window with a fixed window width and sliding step size in the histogram; use the sliding window to traverse the histogram starting from the leftmost side and following the sliding step size; calculate the sum of frequencies within the sliding window before each slide; and store the position of the sliding window and the corresponding sum of frequencies in the frequency comparison library. 3) After the traversal is completed, find the sliding window index and corresponding sliding window position of the maximum frequency of each cluster in the frequency comparison library, and use the frequency value in the sliding window at that position to perform a weighted sum of the pulse repetition frequencies in the window as the representative pulse repetition frequency of each cluster.

5. The signal sorting and identification method under complex electromagnetic environments according to claim 4, characterized in that, The formula for calculating the pulse repetition frequency is: For the first The first time in the film The cluster represents the pulse repetition frequency; In the first The first time in the film The first sliding window of the cluster k Weighted summation of each sliding step; For the first The first time in the film The first sliding window of the cluster k The window covers the first time during the second slide. pulse repetition frequency, The frequency corresponding to the time difference. For the first The first time in the film The first sliding window of the cluster k The total frequency covered by the sliding window range.

6. The signal sorting and identification method under complex electromagnetic environments according to claim 3, characterized in that, In step S3-3, the formula for calculating similarity is: Indicates the first Cluster relative to the first Cluster similarity, For the first The pulse repetition frequency of the cluster, For the first The pulse repetition frequency of the cluster; When the similarity value is less than the set threshold At that time, the first All pulses of the cluster are merged into the first Go into the cluster.

7. The signal sorting and identification method under complex electromagnetic environments according to claim 3, characterized in that, The process of re-merging the second electromagnetic signal cluster in step S4 includes: Step S4-1: Arrange the pulses of each cluster in the second electromagnetic signal cluster in ascending order of pulse arrival time, determine the continuity of pulses in each cluster, and after finding a cluster with a breakpoint, split the cluster from the breakpoint to form two clusters, thus obtaining an electromagnetic signal cluster with continuous time within the cluster. Step S4-2: Calculate the pulse repetition frequency and representative pulse repetition frequency of the cluster head pulse and the cluster tail pulse of each cluster in the time-continuous electromagnetic signal cluster within the cluster. Step S4-3: Calculate the arrival time difference between the tail pulse of each cluster and the head pulse of all other clusters; based on the time difference, merge the two clusters that meet the merging tolerance requirements into one cluster in sequence; after the cluster merging, the third electromagnetic signal cluster is obtained.

8. The signal sorting and identification method under complex electromagnetic environments according to claim 7, characterized in that, The tolerance conditions for the difference are: in, The arrival time difference is the pulse tail of cluster a and the pulse head of cluster b. Let be the representative pulse repetition frequency of the cluster tail pulse of the a-th cluster; Let be the representative pulse repetition frequency of the cluster head pulse of the b-th cluster; This is the tolerance for the difference; and Each of the following is an integer from 0 to 5, and none of them are both 0; when and If any combination of values ​​satisfies the difference tolerance condition, then the cluster tail of cluster a and the cluster head of cluster b are connected and merged into one cluster.

9. The signal sorting and identification method under complex electromagnetic environments according to claim 1, characterized in that, When training the random forest model used in target recognition, the CART decision tree algorithm is used to construct decision trees. When constructing each decision tree, a subset is randomly sampled from the training data for training to ensure the diversity between decision trees. When using the random forest model for prediction, the electromagnetic pulse descriptor sequence read in real time is processed through steps S1-S5 to obtain the fourth electromagnetic signal cluster. The signal data of the fourth electromagnetic signal cluster is then input into the random forest model. Each decision tree in the random forest model completes the predefined classification task based on the input attribute features and outputs the classification result of each cluster. Finally, the target type of each cluster is obtained through voting ensemble. To achieve the target recognition task.

Images