A deep clustering radar pulse signal sorting method based on inner product distance metric

Abstract

The present invention provides a deep clustering radar pulse signal sorting method based on inner product distance metric, which relates to the field of signal and information processing technology. The method first uses a machine learning algorithm for metric learning, utilizes a neural network model based on an attention mechanism to extract features from input data, and then performs a clustering operation using the extracted features. The method uses labeled data for supervised distance metric learning, and simultaneously uses an inner product-based distance metric function and a loss metric function to ensure that the clustering feature distance is consistent with the loss function during training. This method achieves high-accuracy separation of radar signals in complex electromagnetic environments. The algorithm can also detect unknown radars and achieve a certain sorting effect.

Description

Technical Field

[0001] The present invention relates to the technical field of signals and signal processing, and in particular to a deep clustering radar pulse signal sorting method based on inner product distance metric. Background Art

[0002] Signal separation technology is a specialized application of signal processing in the field of reconnaissance and forms the basis for subsequent tasks such as signal identification and jamming. The signal receiver receives interleaved electromagnetic signals, a mixture of interfering signals and pulse signals from multiple radar sources of interest, making signal analysis extremely difficult. Signal separation is the process of extracting the pulse signals from different sources to obtain the pulse signal of a single source. Typically, signal processing involves extracting pulse sequence features from radar intermediate frequency signals and then separating the interleaved pulse features. Signal separation also involves deinterleaving the sequence, also known as signal sorting.

[0003] Clustering methods in the field of signal sorting are usually unsupervised learning methods. These methods use unlabeled sample data, define distance metrics, and utilize the distinguishability of data features to treat samples with close distances as the same class and samples with distant distances as different classes. Radar pulse signals from different radiation sources are classified into different categories, thereby completing radar signal sorting.

[0004] However, distance estimation based solely on the inherent properties of data features is difficult to apply to complex data feature methods. For example, a typical five-dimensional pulse descriptor word (PDW) features include time of arrival (TOA), carrier frequency (RF), pulse width (PW), and angle of arrival (AOA). Within a short period of time, without agility, if the RFs of multiple PDW features are within tolerance, these PDWs generally belong to the same source. However, if the RFs of these pulses are not within tolerance, but the TOA exhibits asymptotic variations, these pulses are likely to belong to the same source. Designing distance metrics for these phenomena requires extensive expert knowledge and rules. In practice, these phenomena are complex and often dependent on the environment, resulting in lengthy, cumbersome, and difficult-to-maintain algorithms.

[0005] Therefore, the present invention proposes a deep clustering radar pulse signal sorting method based on inner product distance metric to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a deep clustering radar pulse signal sorting method based on inner product distance metric. By designing a learning method based on inner product distance metric, radar pulses are expressed using high-dimensional features with high separability, so that the clustering method based on distance metric can be easily used to complete the signal sorting work.

[0007] The present invention is a deep clustering radar pulse signal sorting method based on inner product distance metric, comprising the following steps:

[0008] For the sequence feature X of the radar pulse signal i , mapped to the distance metric space through the sequence neural network model based on the attention mechanism, and obtained the new sequence feature Z i ;

[0009] The parameters of the sequence neural network model based on the attention mechanism are trained to transform the new sequence feature Z i Input the fully connected neural network layer to obtain the logical sub-U i,l , and then use the Softmax function to generate the probability P i,l , and calculate the cross entropy loss function with the label g(i), and then use the momentum stochastic gradient descent algorithm to update all parameters of the model;

[0010] Clustering algorithm based on inner product distance metric for new sequence feature Z i Carry out sorting.

[0011] Furthermore, the sequence neural network model based on the attention mechanism takes the sequence feature X i Map to the distance metric space to obtain the new sequence feature Z i , including three steps, namely the input mapping part, the attention part and the output part, specifically:

[0012] Step 1: Input mapping part, for sequence feature X i ={x i,1 ,x i,2 ,x i,3 …x i,k}, i = 1, 2, 3 ... n, where k is the dimension of the feature, n is the number of pulse data samples sampled from the electromagnetic signal over a period of time, and the sequence feature X is transformed into i Map to the new feature space and obtain features

[0013] Step 2: Attention part, using features Perform similarity analysis and feature fusion to obtain new features

[0014] Step 3: Output part, for new features Similar to step 1, the new features are transformed into Map to the new feature space and output the new sequence feature Z i .

[0015] Furthermore, the 1-D convolutional neural network includes one or more convolutional modules, multiple such convolutional modules are connected sequentially, the output of one module is used as the input of the next module, and the convolutional module is composed of a convolutional layer-ReLU activation function-batch normalization layer. The convolution kernel size of the 1-D convolutional neural network is 1, the output feature dimension is d, and d>k.

[0016] Furthermore, using features Perform similarity analysis and feature fusion, specifically including the following steps:

[0017] Step 21, calculate the sequence feature X i Each data and sequence feature X i All other features in X j , the correlation coefficient S of j≠i i,j :

[0018]

[0019] Step 22: Use the Softmax function to normalize and calculate the calibration coefficient A i,j

[0020]

[0021] Among them, exp() represents the e exponential function;

[0022] Step 23: Weighted sum operation to obtain new features

[0023]

[0024] Furthermore, the new sequence feature Z i Input the fully connected neural network layer to obtain the logical sub-U i,l , the specific steps are:

[0025] Step 4: Introduce a set of learnable parameters W l ∈R d , l=1,2,3…c represents the cluster center, by calculating the feature Z i With W l The inner product of is used as the distance metric, and the distance is used as the criterion for judging the radar pulse feature category. The specific formula is:

[0026]

[0027] Then use the Softmax function to normalize and generate the probability P i,l ,

[0028]

[0029] Among them, exp() represents the e exponential function, P i,l It's Z i The probability of belonging to the lth category, finally, Z i By P i,l ,l=1,2,3…c is determined by the category corresponding to the maximum value;

[0030] The cross entropy loss function is calculated using the label g(i),

[0031]

[0032] Among them, q i,j True probability:

[0033]

[0034] Among them, g(i) is the data category label.

[0035] Furthermore, in the model parameter training stage, the momentum stochastic gradient descent method is used to optimize the cross entropy loss function, update all model parameters, and use a single sample for one update, with the learning rate set to 0.0001.

[0036] Furthermore, the clustering algorithm is a Kmeans clustering algorithm, a hierarchical clustering algorithm or a KNN algorithm.

[0037] Furthermore, the clustering algorithm based on the inner product distance metric is used to classify the feature sequence Z i To sort, specifically use the Kmeans algorithm to sort the feature sequence Z i For sorting, the steps are as follows:

[0038] Step 51, randomly in the sequence feature Z i Select K samples as the center, denoted as V m ,m=1,2,3…K;

[0039] Step 52, use the inner product to calculate the sequence feature Z i to V m distance

[0040] Step 53: Select the center category with the smallest distance as the sequence feature Z i Category,

[0041] c i =argmaxm D i,m ;

[0042] Step 54: average all features of the same category to obtain a new category center:

[0043]

[0044]

[0045] Step 55, repeat steps 52-54, the number of iterations is t, and the final output category c i As a result of signal sorting.

[0046] The beneficial effects of the present invention are as follows:

[0047] 1. The present invention is a deep clustering radar pulse signal sorting method based on the inner product distance metric. Signal sorting is accomplished through Kmeans clustering. This method differs from previous methods primarily in that it uses a sequential neural network model based on an attention mechanism for feature extraction, significantly increasing the separability of data features. Through supervised learning using labeled data, the model can autonomously learn the features required for sorting, reducing the workload of manual feature extraction. Furthermore, this method uses inner product clustering to align the loss function in the model training phase with the distance metric in the clustering phase, ensuring the effectiveness and accuracy of sorting. Highly accurate sorting provides strong support for subsequent signal recognition modules and algorithms, enhancing detection efficiency.

[0048] 2. This invention is a deep clustering radar pulse signal sorting method based on the inner product distance metric. It proposes a sequential neural network model based on an attention mechanism, which effectively extracts high-dimensional features and fully utilizes contextual information. Compared with previous feature extraction methods based on single pulses, it has the ability to extract sequence features, resulting in more robust extracted information. The high-accuracy sorting framework designed by this invention is conducive to enhancing the capabilities of subsequent modules such as model recognition and intent recognition, thereby improving detection capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort, among which:

[0050] Figure 1 This is a structural diagram of a sequential neural network model of the attention mechanism of the present invention;

[0051] Figure 2 Schematic diagram of the clustering method of the present invention. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present invention and are not intended to limit the present invention. That is, the embodiments described herein are only some embodiments of the present invention, not all embodiments. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in various different configurations.

[0053] It should be noted that the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, product or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, product or apparatus.

[0054] The present invention proposes a deep clustering radar pulse signal sorting method based on inner product distance metric.

[0055] The present invention is divided into two parts: one is the sequential neural network model training and inference part based on the attention mechanism, the main purpose of which is to extract high-dimensional features; the other is the clustering algorithm part based on the inner product distance metric, which completes the data (high-dimensional feature representation) sorting task.

[0056] The first is the feature extraction part. For the data X using the radar pulse signal feature method (represented by PDW), i ={x i,1 ,x i,2 ,x i,3 …x i,k}, i = 1, 2, 3 ... n, where k is the dimension of the feature and n is the number of pulse data samples (such as the number of PDWs) sampled from the electromagnetic signal over a period of time. i , it can be mapped to the distance metric space through the sequence neural network model to obtain the new sequence feature Z i .

[0057] The present invention mainly uses a sequence neural network model based on the attention mechanism, and the model structure is as follows: Figure 1 As shown, the framework of the present invention is also compatible with other sequence models, such as RNN, LSTM and other models. At the same time, non-sequential convolutional neural network models can also be used in non-pipeline scenarios. The sequence neural network model based on the attention mechanism calculates each data feature X iand all other features X over a period of time j ,j≠i to determine the similarity, and by merging similar features, similar X i and X j The features of X are becoming more and more similar, and the distant X i and X j Specifically, a module of the attention-based sequential neural network model consists of three steps: input mapping, attention, and output.

[0058] Specifically:

[0059] Step 1: Input mapping part, for sequence feature X i ={x i,1 ,x i,2 ,x i,3 …x i,k}, i = 1, 2, 3 ... n, where k is the dimension of the feature, n is the number of pulse data samples sampled from the electromagnetic signal over a period of time, and the sequence feature X is transformed into i Map to the new feature space and obtain features

[0060] Step 2: Attention part, using features Perform similarity analysis and feature fusion to obtain new features

[0061] Step 3: Output part, for new features Similar to step 1, the new features are transformed into Map to the new feature space and output the new sequence feature Z i .

[0062] For the new sequence feature Z output i ,The present invention uses supervised learning to constrain it so that the features of pulse mapping of the same radiation source are similar, and the features of pulse mapping of different radiation sources are far apart.

[0063] Using the learnable parameter W l The clustering process (that is, the fully connected layer) is used to train the sequence neural network model and update the parameters of the model.

[0064] The clustering-based sorting part needs to use the high-dimensional feature Z obtained in the above feature extraction part i, with the property that the same type of signals are close and different type of signals are far away, by defining the distance metric and utilizing the distinguishability of the data features themselves, the samples with close distances are regarded as the same type and the samples with far distances are regarded as different types, and the radar pulse signals of different radiation sources are classified into different categories, thereby completing the radar signal sorting, such as Figure 2 shown.

[0065] The present invention uses the Kmeans clustering method, but all distance-based clustering methods can be used within the framework of the present invention, such as hierarchical clustering, and some disguised distance-based proximity methods can also be used, such as KNN.

[0066] Example 1

[0067] This implementation example uses PDW input as an example, focusing on the sequential neural network model structure of the attention mechanism used and the size of the intermediate data. The specific implementation steps are as follows:

[0068] S1, input feature X, the dimension is 1000×5, where 1000 is the number of samples in the batch and 5 is the dimension of PDW feature;

[0069] S2. Extract new features Z from feature X using a sequential neural network model based on an attention mechanism. The extracted feature dimension is 1000×512.

[0070] S2.1. Input feature X into the 1-D convolutional neural network model. The model has three layers. The first layer is the convolution layer (the convolution kernel size is 1, the step size is 1, and the output feature dimension is 512). The second layer is the ReLU activation function. The third layer is the batch normalization layer (BatchNormalize, eps is set to 0.0001). The output feature is recorded as Y in , feature dimension 1000×512;

[0071] S2.2, Y in Transpose to obtain Fast calculations using matrix multiplication The output similarity parameter is S and the size is 1000×1000;

[0072] S2.3. Normalize the similarity S in the second dimension using the Softmax function to obtain the calibration coefficient A with a size of 1000×1000.

[0073] S2.4. Take out one row of the calibration coefficient A to obtain Ai, with a size of 1000. Then expand the dimension and copy it to obtain Ai′=[Ai,Ai……,Ai], where Ai is a column vector and the dimension of the matrix Ai′ is 1000*512. Use dot multiplication with the transposed feature to perform fast calculations. Get the weight matrix The dimension is 1000*512, and then the weighted matrix is ​​summed along the first dimension to obtain the feature Perform the same operation on all row vectors Ai of the calibration coefficient matrix A, and obtain the features Perform splicing to obtain the feature matrix Y att , with a size of 1000×512.

[0074] S3, feature Y att Input 1-D convolutional neural network model, which has three layers. The first layer is convolution layer (convolution kernel size is 1, stride is 1, output feature dimension is 512), the second layer is ReLU activation function, and the third layer is batch normalization layer (BatchNormalize, eps is set to 0.0001). The new feature output is denoted as Z, and the feature dimension is 1000×512.

[0075] S4. (During training) input the new feature Z into the fully connected layer (output dimension is 3) to obtain the logical sub-sub ...

[0076] S5. (Inference) Update the new feature Z using the Kmeans algorithm;

[0077] S5.1. Randomly select 3 samples from feature Z as the center category, denoted as V, with a size of 3×512;

[0078] S5.2. Use the inner product to calculate the distance D between feature Z and V, with a size of 1000 × 3;

[0079] S5.3. Select the center category with the smallest distance as the feature category C, with a size of 1000;

[0080] S5.4. Average all features of the same category to obtain a new category center V with a size of 3×512.

[0081] S5.5. Repeat the steps S5.2-S5.4 100 times.

[0082] The final output category C is the result of sorting.

[0083] This invention replaces manual distance measurement with a machine learning-based approach. By using labeled data for supervised learning and employing a feature mapping method based on a sequential neural network model with an attention mechanism, the data is mapped from the radar pulse feature space to the distance metric space. This makes the data more distinguishable in this new space, with similar signals being closer and different signals being farther apart. Furthermore, radar pulses are sorted using distance metric-based clustering methods, such as KMeans.

[0084] The above description is merely 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 conceived by a person skilled in the art within the technical scope disclosed by the present invention without inventive effort should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined in the claims.

Claims

1. A deep clustering radar pulse signal sorting method based on inner product distance metric, characterized in that: The following steps are involved: Sequence characteristics of radar pulse signals , mapped to the distance metric space through the sequence neural network model based on the attention mechanism to obtain new sequence features ; Parameter training of the sequence neural network model based on the attention mechanism is carried out to integrate the new sequence features Input the fully connected neural network layer to obtain the logical sub , and then use the Softmax function to generate the probability , and with the label Calculate the cross entropy loss function and then use the momentum stochastic gradient descent algorithm to update all parameters of the model; Clustering algorithm based on inner product distance metric for new sequence features Sorting is performed; the inner product distance is the inner product of the feature and the learnable parameter vector.

2. A deep clustering radar pulse signal sorting method based on inner product distance metric according to claim 1, characterized in that: The sequence neural network model based on the attention mechanism combines sequence features Map to the distance metric space to obtain new sequence features , including three steps, namely the input mapping part, the attention part and the output part, specifically: Step 1: Input mapping part, for sequence features ,in k is the dimension of the feature, n is the number of pulse data samples sampled from the electromagnetic signal over a period of time, and the sequence features are transformed into Map to the new feature space and obtain features ; Step 2: Attention part, using features Perform similarity analysis and feature fusion to obtain new features ; Step 3: Output part, for new features , similar to step 1, the new features are transformed into Map to a new feature space and output new sequence features .

3. A deep clustering radar pulse signal sorting method based on inner product distance metric according to claim 2, characterized in that: The 1-D convolutional neural network includes one or more convolutional modules. Multiple such convolutional modules are connected in sequence. The output of one module is used as the input of the next module. The convolutional module consists of a convolutional layer-ReLU activation function-batch normalization layer. The convolution kernel size of the 1-D convolutional neural network is 1, the output feature dimension is d, and .

4. The method for deep clustering radar pulse signal sorting based on inner product distance metric according to claim 2, wherein: Using Features Perform similarity analysis and feature fusion, specifically including the following steps: Step 21, calculate sequence features Each data and sequence feature All other features Correlation coefficient : Step 22: Normalize using the Softmax function and calculate the calibration coefficient in, represents the e-exponential function; Step 23: Weighted sum operation to obtain new features : 。 5. The method for deep clustering radar pulse signal sorting based on inner product distance metric according to claim 1, wherein: The new sequence features Input the fully connected neural network layer to obtain the logical sub , the specific steps are: Step 4: Introduce a set of learnable parameters Represents the cluster center, by calculating the features and The inner product of is used as the distance metric, and the distance is used as the criterion for judging the radar pulse feature category. The specific formula is: Then use the Softmax function to normalize and generate the probability , in, represents the e-exponential function, yes Belong to l The probability of each category, finally, Depend on The category corresponding to the maximum value is determined; Use tags Calculate the cross entropy loss function, in, True probability: in, is the data category label.

6. The method for deep clustering radar pulse signal sorting based on inner product distance metric according to claim 5, characterized in that: During the model parameter training phase, the momentum stochastic gradient descent method is used to optimize the cross entropy loss function, update all model parameters, and use a single sample for one update. The learning rate is set to 0.0001.

7. The method for deep clustering radar pulse signal sorting based on inner product distance metric according to claim 1, characterized in that: The clustering algorithm is a Kmeans clustering algorithm, a hierarchical clustering algorithm or a KNN algorithm.

8. The method for deep clustering radar pulse signal sorting based on inner product distance metric according to claim 7, characterized in that: Clustering algorithm based on inner product distance metric for feature sequence To sort, specifically use the Kmeans algorithm to sort the feature sequence For sorting, the steps are as follows: Step 51, randomly in the sequence feature Select K samples as the center category, denoted as Step 52, use inner product to calculate sequence features arrive distance ; Step 53: Select the center category with the smallest distance as the sequence feature Category, ； Step 54: average all features of the same category to obtain a new category center: Step 55, repeat steps 52-54, the number of iterations is t, and the final output category As a result of signal sorting.

Images