PDW signal sorting method based on depth auto-encoder
By combining deep autoencoding network and DBSCAN clustering model, the accuracy and robustness of signal sorting in complex electromagnetic signal environments are solved, and high accuracy and robust signal sorting in dynamic noise environments are achieved.
Patent Information
- Application Number
- CN202510097212.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-22
- Publication Date
- 2025-05-13
AI Technical Summary
In complex electromagnetic signal environments, it is difficult for the prior art to quickly and accurately sort pulse sequences of different radiation sources from interleaved pulse flows, and the signals are susceptible to noise interference, affecting the robustness of sorting.
A deep self-coding network and DBSCAN clustering composition model is introduced, and the deep embedding features of each sample point in the pulse aliasing sequence are obtained through the encoding output of the deep self-coding network, which improves the sorting accuracy, and uses DBSCAN clustering to effectively sort the pulse aliasing sequences within each beat.
Improve the accuracy and robustness of signal sorting in dynamic noise environments, enhance the generalization performance of the model, reduce sensitivity to external interference, and reduce costs.
Smart Images
Figure CN119989020A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a highly robust signal feature sorting method, belonging to the technical field of electronic reconnaissance and signal processing. Background Art
[0002] The goal of signal sorting is to analyze the received signal to determine the number, direction and other important information of the radiation source. It is an important key technology in modern electronic reconnaissance systems. The current electromagnetic signal environment is complex and changeable. New radars continue to emerge and their anti-interference capabilities are constantly improving and strengthening. Electronic reconnaissance faces a more severe test. Electromagnetic signal sorting technology is required to quickly and accurately sort out pulse sequences of different radiation sources from complex and intertwined pulse streams. It plays a key role in subsequent radiation source identification and situational awareness.
[0003] Signals are easily interfered by various noise sources during transmission, which seriously affects the quality of the signal. Therefore, it is of great significance to study how to enhance the robustness of signal sorting. When performing electromagnetic signal sorting, the introduction of an unsupervised autoencoder model not only enhances the robustness of the signal, but also reduces the interference of the signal sorting system from the outside world, making it more reliable and robust, while also saving costs. Summary of the invention
[0004] Purpose of the invention: In order to improve the robustness and model generalization performance of electromagnetic signal characteristics, the present invention introduces a deep autoencoder network deep algorithm and forms a model with DBSCAN clustering. The deep embedded features of each sample point in the pulse aliasing sequence are obtained through the encoding output of the deep autoencoder network, thereby improving the sorting accuracy in a dynamic noise environment.
[0005] Technical solution of the present invention: To achieve the purpose of the present invention, the technical solution adopted by the present invention is to introduce a deep autoencoder network into the cluster sorting model, which specifically includes the following steps:
[0006] A PDW signal sorting method based on deep autoencoder. The main process of electromagnetic signal sorting and feature learning in complex electromagnetic environment is as follows:
[0007] (1) Data preprocessing: Filter, clean, and convert a large number of data stream aliasing pulses to delete redundant state information and empty data, improve data quality, and obtain data related to PDW information.
[0008] (2) Data dimensionality reduction: Based on the deep autoencoder model, the preprocessed input data stream pulses are subjected to data dimensionality reduction to learn the characteristics of PDW data.
[0009] (3),Feature sorting: According to the received signal pulse sequence, the pulse sequences from the same radiation source are clustered, and the pulse sequences from different radiation sources are distinguished.
[0010] (4) Output: After feature sorting, the data of each feature category is divided according to the radiation source data category to obtain description words belonging to different radiation sources.
[0011] The step (2) is specifically as follows: the aliased pulse of the data stream after preprocessing is five-dimensional data, and the deep autoencoder model is used for dimensionality reduction training. After training, the model outputs the input five-dimensional data as three-dimensional data, so that the PDW signal is more robust. The training model of the deep autoencoder model is as follows:
[0012] The model of a deep autoencoder consists of two parts: an encoder and a decoder. The encoder part consists of multiple stacked hidden layers, each of which is responsible for gradually compressing the input data into a more compact representation. Each hidden layer contains multiple neurons that learn the characteristics of the input data through weights and bias parameters. Layer-by-layer calculations map the input data to a latent space; specifically, an encoder is a function that receives an input x and generates an encoded representation h:
[0013] h=σ(W x ·x+b)
[0014] h is the potential representation part in the deep autoencoder. x represents the weight, b represents the bias parameter, and σ is the activation function of each neuron. There are many options, listed as follows:
[0015] Sigmoid function: The value range of the sigmoid function is between (0,1), which corresponds to the range of probability values, and the output values are all greater than 0. Its calculation formula is:
[0016]
[0017] Hyperbolic tangent function (tanh): The output mean of the tanh function is 0, the range is (-1, 1), and the convergence speed is faster than the sigmoid function. The tanh function is expressed as:
[0018]
[0019] Rectified Linear Unit (ReLU): The ReLU function replaces all negative values with 0 and directly outputs positive numbers, with the characteristic of unilateral suppression:
[0020] ReLU(x)=max(0,x)
[0021] The decoder is a mirror image of the encoder, consisting of multiple stacked hidden layers. The decoder is responsible for mapping the latent representation back to the original input space, restoring the original data as much as possible. The calculation of each layer gradually unzips the compression of the latent representation and restores the structure of the original data, that is,
[0022]
[0023] Among them, W h Represents the weights of the linear layer.
[0024] The potential representation part is the feature extraction result of the original data learned by the deep autoencoder. The overall workflow of the deep autoencoder is a compression and reconstruction process, and the reconstruction error can be used as an indicator to measure its working status. The training of the deep autoencoder aims to minimize the reconstruction error, and its loss function can be written as: Where dist is the distance measurement function between the two, using MSE (mean square error):
[0025]
[0026] The step (3) is specifically as follows: using the DBSCAN clustering method, which defines a cluster as the largest set of density-connected points, can divide areas with sufficiently high density into clusters, and can find clusters of arbitrary shapes in a noisy spatial database.
[0027] Beneficial effects of the present invention: For the continuously arriving data stream aliasing pulse sequence, the effective sorting of the pulse aliasing sequence in each beat can be effectively realized based on DBSCAN clustering. In order to improve the robustness of signal features and the generalization performance of the model, while saving costs, a deep autoencoder network is introduced to realize electromagnetic signal sorting based on a deep network. The model is composed of two modules, a deep autoencoder network and a DBSCAN clustering module. The deep embedded features of each sample point in the pulse aliasing sequence are obtained through the encoding output of the deep autoencoder network, realizing the deep mapping of electromagnetic features, effectively improving the generalization performance of the features, thereby improving the sorting accuracy in a dynamic noise environment, and then realizing sorting based on DBSCAN clustering. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 It is an electromagnetic signal sorting model based on the deep autoencoder DBSCAN.
[0029] Figure 2 This is a comparison chart of the sorting accuracy between the deep autoencoder DBSCAN electromagnetic signal sorting model and the DBSCAN electromagnetic signal sorting model in the experiment of the present invention. DETAILED DESCRIPTION
[0030] The present invention will be further explained below in conjunction with the accompanying drawings.
[0031] The technical solution adopted by the present invention is based on a deep autoencoder and includes the following steps:
[0032] A PDW signal sorting method based on deep autoencoder. The main process of electromagnetic signal sorting and feature learning in complex electromagnetic environment is as follows:
[0033] (1) Data preprocessing: filtering, cleaning, converting, and other operations are performed on a large number of data stream aliasing pulses to delete redundant state information and empty data, improve data quality, and obtain data related to PDW information.
[0034] (2) Data dimensionality reduction: In order to reduce the demand for storage and computing resources and improve the generalization ability of the model, the input preprocessed data stream pulses are subjected to data dimensionality reduction based on the deep autoencoder model to learn the characteristics of PDW data.
[0035] (3) Feature sorting: PDW data feature sorting is an important component of electronic reconnaissance. The sorting results will directly affect subsequent electronic reconnaissance links such as situational awareness, tracking and interference. According to the received signal pulse sequence, the pulse sequences from the same radiation source are clustered, and the pulse sequences from different radiation sources are distinguished.
[0036] (4) Output: Based on the above feature sorting process, the data of each type of feature is divided according to the radiation source data category to obtain the description words belonging to different radiation sources.
[0037] The step (2) is specifically as follows: the aliased pulse of the data stream after preprocessing is five-dimensional data, and the deep autoencoder model is used for dimensionality reduction training. After training, the model outputs the input five-dimensional data as three-dimensional data, so that the PDW signal is more robust. The training model of the deep autoencoder model is as follows:
[0038] The Deep Autoencoder model is a deep learning model that belongs to the category of unsupervised learning. It is often used to learn effective representations of data, especially in the tasks of dimensionality reduction and feature learning. The basic idea of the autoencoder is to learn a compact representation of the data by encoding the input data into a compressed representation and then decoding it back to the original data.
[0039] The model of a deep autoencoder consists of two parts: an encoder and a decoder. The encoder part consists of multiple stacked hidden layers, each of which is responsible for gradually compressing the input data into a more compact representation. Each hidden layer contains multiple neurons that learn the characteristics of the input data through weights and bias parameters. Layer-by-layer calculations map the input data to a latent space; specifically, an encoder is a function that receives an input x and generates an encoded representation h:
[0040] h=σ(W x ·x+b)
[0041] h is the potential representation part in the deep autoencoder. x represents the weight, b represents the bias parameter, and σ is the activation function of each neuron. There are many options, listed as follows:
[0042] Sigmoid function: The value range of the sigmoid function is between (0,1), which corresponds to the range of probability values, and the output values are all greater than 0. Its calculation formula is:
[0043]
[0044] Hyperbolic tangent function (tanh): The output mean of the tanh function is 0, the range is (-1, 1), and the convergence speed is faster than the sigmoid function. The tanh function is expressed as:
[0045]
[0046] Rectified Linear Unit (ReLU): The ReLU function replaces all negative values with 0 and directly outputs positive numbers, with the characteristic of unilateral suppression:
[0047] ReLU(x)=max(0,x)
[0048] In this study, RELU was selected as the activation function.
[0049] The decoder is a mirror image of the encoder, consisting of multiple stacked hidden layers. The decoder is responsible for mapping the latent representation back to the original input space, restoring the original data as much as possible. The calculation of each layer gradually unzips the compression of the latent representation and restores the structure of the original data, that is,
[0050]
[0051] Among them, W h Represents the weights of the linear layer.
[0052] It seems that no changes have been made to the original data, but its potential representation is the feature extraction result of the original data learned by the deep autoencoder. Compared with the traditional activation function that uses multiple nonlinear transformations, its multi-layer structure can flexibly capture the nonlinear characteristics of the input data, so it can be used as a nonlinear dimensionality reduction method. The overall workflow of the deep autoencoder is a compression and reconstruction process, and the reconstruction error can be used as an indicator to measure its working status. The training of the deep autoencoder aims to minimize the reconstruction error, and its loss function can be written as: Where dist is the distance measurement function between the two, usually using MSE (mean square error):
[0053]
[0054] A deep autoencoder is essentially a feedforward neural network, so it also benefits from the advantages brought by the network depth like ordinary feedforward neural networks. A deep autoencoder refers to a deep autoencoder that contains at least one hidden layer in the encoding stage. According to the universal approximation theorem, a feedforward neural network with one hidden layer and a sufficient number of hidden layer neurons can represent the approximate values of a large part of the function with arbitrary precision. However, if the layer from input to encoding is shallow, that is, there is no hidden layer in the middle, it is impossible to add arbitrary constraints to it. The deep autoencoder overcomes this limitation, which means that when the number of hidden layer neurons is large enough, the mapping from input data to low-dimensional encoding can be approximated with arbitrary precision.
[0055] The step (3) is specifically as follows: using the DBSCAN clustering method, which defines a cluster as the largest set of density-connected points, can divide areas with sufficiently high density into clusters, and can find clusters of arbitrary shapes in a noisy spatial database.
[0056] The specific process of the DBSCAN algorithm is as follows:
[0057] Input: Dataset, given the minimum number of neighborhood points to become a core object in the neighborhood: MinPts, neighborhood radius: Eps;
[0058] Output: cluster set;
[0059]
[0060] First, the pulse PDW time series data of 30 radiation sources were simulated and generated. The above radiation sources were deployed in a specific scenario, and the radiation source individuals were effectively sorted according to the received pulse aliasing sequence within a fixed step time beat. The pulse aliasing sequence of 30 radiation sources contains a total of 6000 pulse PDW samples, which are divided into 6000 beats, with an average of 10466 samples in each beat.
[0061] To fit the real scene, the pulse data of 30 radiation sources reflect the different modulation characteristics of the radiation sources in various dimensional parameters, such as repetition mode, including fixed repetition, staggered repetition, sliding repetition, etc.; the carrier frequency has agile frequency characteristics. Different radiation sources may have similar center values in some dimensions, or have different proportions of noise error values, and the active / silent periods of each radiation source are also different.
[0062] The proposed DBSCAN clustering radiation source sorting method based on deep autoencoder is verified by simulation experiments. The algorithm performance is evaluated by clustering accuracy, false alarm and other indicators. Clustering accuracy is calculated based on the ratio of pulses of the same radiation source in a specific beat to be clustered into the same cluster; false alarm means that the pulses of the same radiation source are clustered into 2 or more clusters, that is, the number of clusters is greater than the number of real radiation sources in the beat.
[0063] First, the aliased pulse sorting performance of DBSCAN clustering is tested. For the pulse aliasing sequences of 30 radiation sources, according to the statistical characteristics of the sample distribution in each dimension, the arrival angle, pulse width and center frequency parameters in the pulse description word PDW are used for clustering, and the eps parameter of the DBSCAN clustering algorithm is set to 0.3.
[0064] Similar to the above process, the pulse aliasing sequence of 30 radiation sources is sorted by the sorting algorithm based on deep DBSCAN clustering. The arrival angle, pulse width and center frequency parameters in the pulse description word PDW are also selected for clustering. The eps parameter setting of the clustering algorithm is also set to 0.3, and the hidden variable dimension of the deep autoencoder network encoding output is set to 12. Finally, the sorting accuracy based on deep autoencoder DBSCAN clustering and DBSCAN clustering is obtained. Through simulation experiments, the average clustering accuracy based on deep DBSCAN clustering reaches 98.6%. It can also be concluded that by introducing deep autoencoders in DBSCAN clustering, the generalization of extracted features is improved, and the reliability and robustness of data feature sorting are enhanced, which is less disturbed by the outside world and more robust.
Claims
1. A PDW signal sorting method based on deep autoencoder, characterized in that: Here are the steps: (1) Data preprocessing: filtering, cleaning, and converting a large number of aliased pulses in data streams to delete redundant state information and empty data, improve data quality, and obtain data related to PDW information; (2) Data dimensionality reduction: Based on the deep autoencoder model, the input preprocessed data stream pulses are subjected to data dimensionality reduction to learn the characteristics of PDW data; (3),Feature sorting: according to the received signal pulse sequence, the pulse sequences from the same radiation source are clustered, and the pulse sequences from different radiation sources are distinguished; (4) Output: After feature sorting, the data of each feature category is divided according to the radiation source data category to obtain description words belonging to different radiation sources.
2. The PDW signal sorting method based on deep autoencoder according to claim 1 is characterized in that: The step (2) is specifically as follows: the aliased pulse of the data stream after preprocessing is five-dimensional data, and the deep autoencoder model is used for dimensionality reduction training. After training, the model outputs the input five-dimensional data as three-dimensional data, so that the PDW signal is more robust; the training model of the deep autoencoder model is as follows: The model of a deep autoencoder consists of two parts: an encoder and a decoder. The encoder part consists of multiple stacked hidden layers, each of which is responsible for gradually compressing the input data into a more compact representation. Each hidden layer contains multiple neurons, which learn the characteristics of the input data through weights and bias parameters. Layer-by-layer calculations map the input data to a latent space. Specifically, the encoder is a function that receives an input x and generates an encoded representation h: h=σ(W x ·x+b) h is the potential representation part in the deep autoencoder; where W x represents the weight, b represents the bias parameter, and σ is the activation function of each neuron. There are many options, listed as follows: Sigmoid function: The value range of the sigmoid function is between (0,1), which corresponds to the range of probability values, and the output values are all greater than 0; its calculation formula is: Hyperbolic tangent function: The output mean of the tanh function is 0, the range is (-1, 1), and the convergence speed is faster than the sigmoid function; the tanh function is expressed in the form of: Rectified Linear Unit: The ReLU function replaces all negative values with 0 and directly outputs positive numbers, which has the characteristics of unilateral inhibition: ReLU(x)=max(0,x) The decoder is a mirror image of the encoder, consisting of multiple stacked hidden layers. The decoder is responsible for mapping the latent representation back to the original input space, restoring the original data as much as possible. The calculation of each layer gradually unzips the compression of the latent representation and restores the structure of the original data, that is, Among them, W h represents the weight of the linear layer; The potential representation part is the feature extraction result of the original data learned by the deep autoencoder; the overall workflow of the deep autoencoder is a compression and reconstruction process, and the reconstruction error can be used as an indicator to measure its working status; the training of the deep autoencoder aims to minimize the reconstruction error, and its loss function can be written as: Where dist is the distance metric function between the two, using MSE:
3. The PDW signal sorting method based on deep autoencoder according to claim 1 is characterized in that: The step (2) is specifically as follows: using the DBSCAN clustering method, which defines a cluster as the largest set of density-connected points, can divide areas with sufficiently high density into clusters, and can find clusters of arbitrary shapes in a noisy spatial database.