CN113381827B

CN113381827B - Electromagnetic signal spectrum sensing method based on deep clustering network

Info

Publication number: CN113381827B
Application number: CN202110640928.6A
Authority: CN
Inventors: 臧彧; 李嘉廉; 王强; 王程; 陈修桥; 车吉斌
Original assignee: Xiamen University
Current assignee: Xiamen University
Priority date: 2021-06-09
Filing date: 2021-06-09
Publication date: 2022-04-26
Anticipated expiration: 2041-06-09
Also published as: CN113381827A

Abstract

The invention discloses an electromagnetic signal spectrum sensing method based on a deep clustering network, which comprises the following steps: s1, sorting and positioning samples according to the input signals, and carrying out regularization processing on the radiation source signals; s2, clustering the input radiation source signals by adopting a deep continuous clustering network; s3, utilizing the clustered radiation source signals, combining the input positioning results, and using a target tracking algorithm to realize the front-back association of a plurality of time signals to form the motion trail of the electronic target; s4, analyzing the motion rules among a plurality of electronic targets according to the motion tracks of the electronic targets, and adopting a set updating algorithm and a platform attachment algorithm to carry out platform aggregation to discover potential target platforms; according to the invention, by inputting the sorted and positioned signal samples, the system can automatically distinguish signals from different devices, and can perform subsequent electronic target association and platform aggregation work according to the device information.

Description

Electromagnetic signal spectrum sensing method based on deep clustering network

Technical Field

The invention relates to the technical field of electromagnetic signal identification, in particular to an electromagnetic signal frequency spectrum sensing method based on a deep clustering network.

Background

Electromagnetic signal spectrum sensing is a major research direction in the field of electronic countermeasures. In the modern information war, corresponding countermeasures can be taken in time only by correctly analyzing the situation of enemy radar signals, and the efficiency of weapons and forces of our party is fully exerted.

In equipment identification based on radiation source signals, a traditional identification algorithm relies on a radar knowledge base which is manually maintained, and after the signals are captured, parameter comparison is carried out on the signals one by one with radar signals in the knowledge base. However, such methods still rely to some extent on human intervention. And is not motivated by the emerging radar signal beam. Although classical clustering algorithms can distinguish unknown radar signals without a priori knowledge base. However, when the dimension of the input signal is increased, the calculation of the distance on the sparse high-dimensional features will result in the loss of information, and therefore an additional feature selection process or dimension reduction is required. However, the process of classical clustering is discrete and cannot be optimized together with the input coding part, which also results in poor algorithm performance.

Disclosure of Invention

The invention aims to provide an electromagnetic signal spectrum sensing method based on a deep clustering network, which is characterized in that signals from different devices are automatically distinguished by a system through inputting sorted and positioned signal samples, and subsequent electronic target association and platform aggregation can be carried out according to the device information.

In order to achieve the purpose, the invention adopts the following technical scheme:

an electromagnetic signal spectrum sensing method based on a deep clustering network comprises the following steps:

s1, sorting and positioning samples according to the input signals, and carrying out regularization processing on the radiation source signals;

s2, clustering the input radiation source signals by adopting a deep continuous clustering network;

s3, utilizing the clustered radiation source signals, combining the input positioning results, and using a target tracking algorithm to realize the front-back association of a plurality of time signals to form the motion trail of the electronic target;

s4, analyzing the motion rules among a plurality of electronic targets according to the motion tracks of the electronic targets, and adopting a set updating algorithm and a platform attachment algorithm to carry out platform aggregation to discover potential target platforms.

Preferably, the specific process of step S1 is:

s11, sorting and positioning the sample X ═ X according to the input signal₁,X₂,......Xn]Wherein

S12, screening out carrier frequency, repetition frequency and pulse width as the basis for identifying the radiation source signal;

s13, splicing the pulse parameters of the input signal, such as repetition frequency, carrier frequency and pulse width, into a one-dimensional vector;

s14, eliminating the dimension between the different types of parameters by using normalization, wherein the normalization formula is as follows:

wherein x is_i,jFor the ith sample X_iThe jth component of (a), x_i.,jIs the set of jth components of all samples in the input data set.

Preferably, the deep continuous clustering network in step S2 includes a stacked denoising auto-encoder SDAE and a joint loss function that can be optimized using gradient descent to distinguish different types of radiation source signals without a priori knowledge base.

Preferably, the clustering expression of the stacked denoising auto-encoder SDAE is:

wherein X is an input data set;

representing the mapping of a high-dimensional input data set to a low-dimensional space,

representing the use of a low-dimensional reconstruction of a high-dimensional input, F_θAnd G_ωThe method is realized by an encoder and a decoder of a stack denoising automatic encoder respectively;

represents a Frobeneus paradigm.

Preferably, the processing procedure of the joint loss function is as follows: employing proxy collections

The following objective function is optimized:

wherein the content of the first and second substances,

and

all are M-estimation functions, which are used for pulling potential same-class agents to the same point and isolating false links among classes; ε is the full connection graph between X; after the algorithm is finished, a graph is built on Z

And if the distance between the two nodes is smaller than a given threshold value, the two nodes are connected with each other, and each connected component in the finally obtained graph belongs to one class.

Preferably, the specific process of the target tracking algorithm in step S3 is as follows:

s31, predicting the state of each target in the next frame by using a linear constant speed model, wherein the state of each target is shown as the following formula:

wherein u and v represent longitude and latitude coordinates of the object, respectively,

and

respectively representing the speed of the target in latitude and longitude;

s32, establishing a bipartite graph between a predicted target and an actually detected target, wherein the weight is the absolute distance between the targets, a longest displacement needs to be estimated according to the time difference and the maximum movement speed before and after, and target pairs with overlong distance are ignored; after the graph building is finished, solving the optimal weight matching of the bipartite graph by using the minimum cost maximum flow;

and S33, updating the detected state by using Kalman filtering to finish the smoothing of the motion trail and the motion trail of the electronic target.

Preferably, the updating process of the set updating algorithm in step S4 is as follows: firstly, defaulting that all targets are positioned on a platform, and then segmenting the targets according to time and position information; if the distance between two targets is greater than the set value at the same moment, the two targets are not on the same platform; the distribution process of the platform assignment algorithm in step S4 is as follows: if the distance between a new target and the targets of the known platform information is smaller than a set value at the current moment, namely the current moment is considered to be on one platform, the new target is bound to the platform to which the target nearest to the new target belongs, otherwise, a platform is newly built for storing the new target.

After the technical scheme is adopted, the invention has the following beneficial effects: firstly, eliminating the dimension of an input signal through regularization, and unifying the input formats of all working modes; based on formatted data, feeding the data into a deep continuous clustering network, and obtaining the model of the logic equipment on the premise of no prior knowledge so as to distinguish the model from the emission signals of other equipment; completing electronic target association through a target tracking algorithm according to the model information carried by the signal; and finally, completing the work of platform aggregation by using a set updating algorithm and a platform attachment algorithm.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is an input visualization of the present invention showing the distribution of the original input signal on a map;

FIG. 3 is an illustration of the present invention showing the network structure of a stacked noise reduction self-encoder in a deep clustering network;

FIG. 4 is an effect diagram of the present invention, showing that a plurality of radar signals of the same type are fused together by an association algorithm, and the types of logical devices provided by the clustering network can be given;

FIG. 5 is an effect diagram of the present invention, which presents the process of mining a potential platform from an associated electronic target and plotting the motion trajectory of a mobile platform;

fig. 6 is a diagram showing the distribution of the platform found from the input signal data according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

As shown in fig. 1, the present invention is mainly divided into 4 steps:

step S1: sorting and locating samples X ═ X according to input signals₁,X₂,......Xn]Wherein

And screening out carrier frequency, repetition frequency and pulse width as the basis for identifying the radiation source signal, and regularizing the parameters according to the working modes of the three pulse parameters. Fig. 2 shows the location distribution of the radar signal used in the present application, which shows the distribution of the original input signal on the map.

Compared with the independent use of the repetition frequency for radar signal sorting, the repetition frequency, the carrier frequency and the pulse width are used for interactive processing, so that the sorting and identifying precision of the radar of the types such as agility, frequency diversity and the like can be improved.

Different types of radars will have corresponding different operating modes on these three parameters, and table 1 is a common parameter type:

TABLE 1

Pulse parameter	Type (B)
		Carrier frequency (RF)	Fixed, agile, and jumping
Pulse Width (PW)	Fixed, multiple pulse widths
		Repetition frequency (PRI)	Fixed, staggered, jittered, hopped, group-changed

Because the types of the repetition frequency, the carrier frequency and the pulse width parameters of the pulse signals generated by one radar in a certain working mode are basically consistent, the input signals can be primarily screened by utilizing three parameter types.

In the subsequent clustering process, the distance between two input samples needs to be calculated, and in order to enable indexes of different units or magnitudes to be compared and weighted, the dimensions of different indexes need to be eliminated through normalization, and input signals need to be converted into dimensionless expressions. The normalization formula is as follows:

As shown in table 1, different parameters have different types and also present different expressions, which also affect the subsequent clustering algorithm. The manifestations can be divided into the following three types:

fixing: the fixed values, the dispersion values and the grouping variables in the table 1 belong to the type, and the parameters in the sorting signals are stable and consistent in number and fluctuate within a small range of the fixed values. Therefore, whether the parameters are consistent or not is judged, and the distance is judged on the basis of judging whether the number is equal or not. This class can be fed into the clustering algorithm without any modification. It should be noted here that since the input data is a sorted signal, the spread and grouping repeats one or more cycles in a very short time, and therefore all possible values are covered, and thus classified as such.

The interval type: the agility and jitter in table 1 are of this type, which is shown by the fact that the number of parameters in the sorted signal is not uniform, but the occurring values fluctuate within a certain range. If clustering is performed by directly determining the distance, a very large error is caused. The data for such parameters will be normalized to the minimum and maximum values of the recording interval, considering that the range of this interval is approximately uniform during sorting.

Set type: the jumps in table 1 are of this type, which represent the number of parameters in the sorted signal, but the possible values of occurrence are limited. The sorted individual signal corresponds to a parameter that may be only a subset of the parameters for that type of radiation source. Like the interval data, the disparity in number results in the inability to directly calculate the distance. Considering that when other parameters present the same type and numerical characteristics, the value ranges of such parameters are obviously distinguished, and generally no overlapping occurs, so as to achieve the purpose of convenient sorting, a regularization method of a region type can be used for processing the data, and the minimum value and the maximum value of the set are recorded.

In order to quantitatively evaluate the performance of the algorithm, two indexes, namely Accuracy (ACC) and Normalized Mutual Information (NMI), need to be introduced, as shown in the following formula:

wherein, c_iFor genuine tags, here the genuine device type, p_iA predicted value for the clustering algorithm. m is an allocation algorithmAnd establishing a one-to-one mapping relation between C and P. Obviously, ACC builds on the best match of C and P, and this can be solved using bipartite graph weight best match.

In the NMI formula, I represents mutual information, and H represents entropy.

And roughly dividing the test sample into five types according to the parameter types, wherein only three parameters of the first type of radar signals are all fixed. The fixed clustering algorithm was unchanged, the randomized control results were given below, the algorithm was run 10 times, and the results are shown in table 2:

TABLE 2

However, the higher the ACC ↓andnmi ≠ are, the better the display effect is. It can be seen from table 2 that irregular discrete data will cause a huge loss of performance to the clustering algorithm.

Step S2: and clustering the input radiation source signals by adopting a deep continuous clustering network. The network contains a stacked de-noising auto-encoder (SDAE) whose encoder and decoder are each composed of 4 fully connected layers. Extracting low-dimensional representation of an input signal from a pre-trained encoder, performing relaxation operation on the existing discrete clustering loss, and performing joint optimization together with reconstruction loss of the encoder to finally achieve the purpose of distinguishing different types of radiation source signals without a priori knowledge base.

Considering the maximum number of carrier frequencies, repetition frequencies and pulse widths, D needs to be set to 30. The original feature representation is a high-dimensional sparse vector, and the information cannot be fully utilized due to the fact that the distance between features is calculated, so that most clustering algorithms perform poorly on high-dimensional data. It is necessary to project the data into the low dimensional space first

Where D < D, and then performing clustering operations on the low-dimensional space. The low dimensional representation of dataset X is denoted Y ═ Y₁,y₂……y_n]. The function of the projection is noted

Thus, for any i, there is y_i＝f_θ(X_i)。

However, the low-dimensional representation must be constrained to ensure that it can effectively characterize the high-dimensional dataset. Therefore, it is necessary to add a mapping function

The low-dimensional space is projected back to the original high-dimensional space. Finally, an efficient low-dimensional representation is obtained by minimizing the reconstruction loss, namely:

wherein Y is F_θ(X)，Ω＝{θ,ω}，

Represents a Frobeneus paradigm. Where F_θAnd G_ωThe implementation is achieved with an encoder and decoder of a stacked denoising auto-encoder, where both functions are implemented by four fully-connected layers. By F_CRepresenting the fully-connected layer, LeakyRelu representing the nonlinear activation function, F_CThe layers are composed of a linear transformation plus a non-linear activation function, and only the two types of layers contain the parameter weight W and the deviation b to be learned. The structure of the whole network is shown in FIG. 3

The decoder and encoder are fully symmetric network structures and training will take place on each pair of coding and decoding layers in turn, introducing noise by randomly dropping the input. The fully-connected layers of the encoder and decoder will be trained from outside to inside, and when all layers are trained, the entire SDAE continues to fine-tune the overall parameters using the reconstruction loss.

Classical clustering algorithms are discrete, for example KMeans need to calculate the distance of a data point to the centroid at each round and update the centroid distance again; DBScan requires a search method to continuously merge data points with a connected density. The discrete clustering process is not beneficial to the continuous optimization of the low-dimensional feature Y, so that a continuous and optimizable clustering process needs to be designed.

The embodiment employs proxy aggregation

The following objective function is optimized:

wherein the content of the first and second substances,

and

all are M-estimation functions, which are used for pulling potential same-class agents to the same point and isolating false links among classes; epsilon is a full connection graph among X, and in actual use, because the data volume may be very large, a local connection graph can be constructed by kNN, and then clusters are connected by using a minimum spanning tree, so that an approximate full connection graph is obtained. Omega_i,jIs the contribution of each data point to the pair-wise loss:

wherein n is_iIs Z_iDegree in the graph epsilon. λ is used to balance two loss terms, set as

Where A ═ Σ ∑_(i,j)∈εω_i,j(e_i-e_j)(e_i-e_j)^T，||·||²Representing the spectral norm.

Where the first term loss forces the proxies towards the real data point and the second term is essentially a repulsive loss, making the distance between the proxies as far as possible. Optimizing such an objective function has two advantages: firstly, relaxing a discrete clustering process into a fixed continuous objective function; secondly, each data point has a self-agency in Z, and the number of clusters does not need to be set in advance; finally, the estimation function reduces the effect of outliers.

Together with the previous reconstruction loss, the parameters Ω and Z are optimized using such an objective function:

to optimize this objective function, two points need to be noted:

1) the parameter Ω is only related to the first reconstruction loss and the parameter Z is only related to the second and third reconstruction losses, but the update rates of these two parameters are different, and in order to eliminate this imbalance, the adaptive algorithm Adam is used instead of SGD.

2) Because the data volume is large, the data needs to be sent into the network in batches for clustering. Thus, data in a batch may be paired up with data outside the batch, which can make it difficult to resolve data points. Thus, an edge in ε may be sampled, and the corresponding sample, i.e., the set of all vertices connected to the sampled edge. At the same time, this results in the first two terms getting extra weight, since the same sample may be sampled multiple times because of the higher degree. Therefore, the objective function needs to be rebalanced as follows:

wherein the content of the first and second substances,

is the ith node in subgraph epsilon_BDegree ofAnd (4) counting.

The stopping judgment principle of the training process is as follows:

after each round of training is finished, a graph is built on Z

Wherein if

F is then_i,j1.δ is a preset threshold value. If less than 0.1% of the edges change, the algorithm stops training and each connected component is a class in the resulting graph.

Table 3 gives the quantitative analysis of the algorithm herein versus the classical clustering algorithms (KMeans + + and DBScan):

TABLE 3

ACC↑	#1	#2	#3	#4	#5
						Ours	0.997	0.816	0.842	0.740	0.751
KMeans++*	0.912	0.794	0.753	0.699	0.652
						DBScan*	0.923	0.727	0.762	0.687	0.641
KMeans++	0.712	0.622	0.701	0.604	0.579
						DBScan	0.758	0.602	0.680	0.538	0.532
NMI↑	#1	#2	#3	#4	#5
						Ours	0.996	0.835	0.806	0.777	0.803
KMeans++*	0.895	0.801	0.768	0.674	0.716
						DBScan*	0.934	0.769	0.781	0.612	0.698
KMeans++	0.721	0.651	0.721	0.537	0.574
						DBScan	0.763	0.587	0.676	0.567	0.593

KMeans + + and DBScan are both calculations of distance directly on the input signal, and KMeans + +, and DBScan are calculations of distance on the low dimensional representation extracted by SDAE.

As can be seen from table 3, from the comparison of the two pairs of classical algorithms, the low-dimensional representation extracted by SDAE more effectively represents the input signal, thereby improving the algorithm performance of the clustering algorithm; in addition, the method for optimizing a continuous clustering loss function by using gradient descent is more natural and compact in combination with the step of low-dimensional feature extraction, and therefore, the method has better effect than the classical algorithm with discrete steps.

Step S3: and utilizing the clustered radiation source signals, combining the input positioning result, and realizing the front-back association of a plurality of time signals by using a target tracking algorithm to form the motion trail of the electronic target.

The input signal carries model information, via S2, which is essentially a number that can be distinguished from the signals transmitted by other devices. And then, step S3 is carried out to realize the discovery and track extraction of the electronic target and the target platform, and the specific steps are as follows:

on the basis of S2, the input signal carries model information which can be distinguished from other signals, the model information can help to distinguish electronic targets which are close to each other in position and even coincide with each other, and the accuracy of the algorithm can be improved while the time consumption of the electronic target association algorithm is reduced.

Is provided with

Is the T (T epsilon [1, T)]) Set of targets detected at each moment, wherein

Is the latitude and longitude coordinate of the ith target. Plus the implicit target state, i.e. the velocity vector of the target, i.e. the complete state of the target:

and

representing the speed of the target in latitude and longitude, respectively.

To establish target association at two moments, two steps are needed, which are respectively:

1) a prediction of a next location of the target;

2) matching of the predicted location and the observed location;

these two steps are solved using kalman filtering and bipartite graph optimal matching, respectively.

The Kalman filtering is built on a hidden Markov model, and the basic dynamic system can be represented by a Markov chain which is built on a linear operator interfered by Gaussian noise. The state of the system can be represented by a vector whose elements are real numbers. With each increase in discrete time, the linear operator acts on the current state, creating a new state and also introducing noise, and control information from known controllers of the system is also added. At the same time, another linear operator disturbed by noise produces a visible output of these implicit states.

The kalman filter model assumes that the true state at time k evolves from time k-1, i.e.:

x_k＝F_kx_k-1+B_ku_k+w_k

wherein, F_kIs acting on x_k-1A state transition model of (1); b is_kIs acting on the controller vector u_kAn input-control model of (1); w_kIs process noise, corresponds to w_k～N(0,Q_k)；Z_kIs to the true state x_kOne measurement of, i.e. Z_k＝H_kx_k+v_k(ii) a Wherein H_kIs an observation model, maps the true state space into an observation space, v_kIs observed noise, in line with v_k～N(0,R_k)。

Initial state and noise per time { x }₀,w₁,...w_k,v₁...v_kAll considered to be independent of each other.

The state of the kalman filter is represented by two variables:

1)

is an estimate of the state at time k;

2)P_k|k-1the error covariance matrix is estimated a posteriori, and the accuracy of the estimated value is measured.

The operation of the kalman filter comprises two phases: and (4) predicting and updating. In the prediction phase, the filter uses the estimate of the last state to make an estimate of the current state. In the update phase, the filter optimizes the predicted value obtained in the prediction phase using the observed value for the current state to obtain a more accurate new estimated value.

The prediction is divided into a prediction state and a prediction estimation covariance matrix, as shown in the following equation

Updating needs to firstly calculate a measurement residual error, a measurement residual error covariance and an optimal Kalman gain:

and then used to update the filter variables x and P:

P_k|k＝(I-K_kH_k)P_k|k-1

observation at time t-1

After Kalman filtering, the method can obtain

In addition we have the observation of the time t

And performing target association of the time t-1 and the time t by using bipartite graph optimal matching. In order to ensure that the real target is not filtered by the associated algorithm as much as possible, the optimal matching of the weight of the bipartite graph is used, and on the basis, the total matching cost is minimized, namely:

min A_ijC_ij，i∈[1，n_t-1]，j∈[1，n_t]

wherein A is_ijE {0, 1}, is the matching condition of each target pair of the front and back frames, C_ijIs composed of

And

here as the matching cost. If the target at the previous moment is mismatched, the corresponding track stops updating, and if the target at the current moment is mismatched, the track is added into the set as a new track.

Finally obtaining T ═ T₁，t₂，...，t_kH, k total traces. Each timeOne track is

Wherein l_iIs a track t_iLength of (f)_iIs a track t_iThe start frame of (2). w is a_i，jIs a track t_iThe target sequence number in the jth frame.

After the track is obtained, the algorithm filters the track as a research object, and the reference rule is as follows:

1) too short a trajectory and its target will be discarded, which will filter out false alarms where the target characteristics are unstable.

2) The average confidence of all targets contained in the trajectory is calculated, and when the average confidence is too low, we will discard that even if there are detections at several consecutive times, the confidence that is generally low (reflected in the measure of the average) can be considered a false alarm.

3) Calculating the dominant motion direction of the object moving from the t-1 frame to the tth frame, i.e. { alpha }¹，α²，...，α^T-1}. Simultaneous calculation of the trajectory t_iDirection of movement, is noted

If the following equation is satisfied,

|α^j-β^j|≤∈_bearingthe track is also deleted.

Fig. 4 shows the electronic target orientation after target association in this embodiment, wherein the text box shows the device tag deeply clustered in step S2.

Step S4: and analyzing the motion rules among a plurality of electronic targets according to the motion tracks of the electronic targets, and carrying out platform aggregation by using a set updating algorithm and a platform attachment algorithm to discover a potential target platform.

Target association is the identification of a single target, and target platform integrated identification is the process of multiple target integrated identification platforms. Through the electronic target track information obtained in step S3, whether a plurality of radiation sources with similar positions form a large target or not is analyzed, so as to find out a potential target platform and obtain the track thereof.

This step involves two algorithms, a set update algorithm and a platform attach algorithm.

The set update algorithm initially defaults to all targets on one platform and then segments these targets based on time and location information. That is, if two targets are separated by a distance greater than the set value at the same time, they are not considered to be on one platform.

First, an existing platform set S is obtained from a database, and when a system is just initialized, S is an empty set. Each element in the set represents a platform, one platform simultaneously comprises a plurality of targets, each platform in the S is traversed, and the target of the first one of the platforms appearing at the current moment is found out. If no object in the table has transmitted a radiation source signal at the current time, the current information may be deemed insufficient to further segment the table. Clustering the target at the current moment according to the position to obtain a position set S at the current moment₁. For a platform in S, after finding the first target appearing at the current time, note down it at S₁Index of (1) is denoted as i. Next traverse other targets in this platform if not present at S₁Or at S₁If the index in (1) is also i, the original platform number is kept unchanged. Otherwise, the target is assigned to a new platform, which is appended to the set of platforms S.

An example analysis of this algorithm is as follows: assume the previous time, { A₁,A₂,A₃,B₁,B₂,B₃,C₁,C₂D, E } on one platform, where D and E have not transmitted a signal at the current time, and so are temporarily stored on the current platform. { A_i}、{B_iAnd { C }_iIndicates that the objects appear at three different locations at the current moment. According to the algorithm described above, these objects will be split into three sets, respectively { A }₁,A₂,A₃,D,E}、{B₁,B₂,B₃And { C }₁,C₂}。

Since the new targets do not have any platform information before, it is necessary to assign platforms to these new targets using a platform assignment algorithm. The distribution principle is that if the distance between a new target and a plurality of targets with known platform information is smaller than a set value at the current moment, namely the current moment is considered to be on one platform, the new target is bound to the platform to which the target nearest to the new target belongs, otherwise, a platform is newly built for storing the new target.

An example analysis of this algorithm is as follows: assume at the current time, { A₁,A₂,B₁,B₂And D, E distances are similar, namely the distances are temporarily set on the same platform. Where D and E are newly emerging targets, where D is a distance B₁Nearest, E distance A₂Recently, then B will be₁The platform of (A) is assigned to D, A₂Is assigned to E.

For the target Tracking algorithm, the present embodiment uses multi-target Tracking (MOTA) to measure the algorithm performance, the MOTA form is as follows:

wherein FN is False Negative, FP is False Positive, IDSW is ID Switch, and GT is the number of real targets. The MOTA considers object matching errors at all moments in tracking, mainly FN, FP and ID Switch. The MOTA gives a very intuitive measure of the performance of the tracker in detecting objects and maintaining trajectories, independent of the accuracy of the estimation of the object position.

As a result, as shown in Table 4, since the electronic targets are closer to each other, the targets are likely to be staggered, resulting in lower MOTA values than the platform tracking

	MOTA(％)
		Target tracking	84.3
Platform tracking	90.1

As shown in fig. 5, the left image is the electronic target association result obtained after step S3, and the right image is the motion trajectory of the target platform obtained after step S4. Fig. 6 shows all target platforms and their motion trajectories of the data used in the present embodiment.

As mentioned above, the invention provides an electromagnetic signal spectrum sensing method based on a deep clustering network. The regularized sorted positioning signal data is fed into a deep clustering network to distinguish radar signals emitted by different devices. The deep clustering network uses a stacking denoising automatic encoder to force the network to learn effective low-dimensional features, and achieves the purpose of clustering by optimizing continuous and differentiable loose clustering loss on a low-dimensional feature space. And combining the logic equipment information given by clustering, combining Kalman filtering and optimal matching of bipartite graph weights, obtaining the motion trail of the electronic target, and further exploring potential comprehensive platform information through a platform aggregation step.

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An electromagnetic signal spectrum sensing method based on a deep clustering network is characterized by comprising the following steps:

the deep continuous clustering network in the step S2 comprises a stacking denoising automatic encoder SDAE and a joint loss function which can use gradient descent optimization, so that different types of radiation source signals can be distinguished without a priori knowledge base;

the clustering expression of the SDAE is as follows:

wherein X is an input data set;

represents a Frobeneus paradigm;

the processing procedure of the joint loss function is as follows: employing proxy collections

The following objective function is optimized:

wherein the content of the first and second substances,

and

If the distance between the two nodes is smaller than a given threshold value, the two nodes are communicated with each other, and each communicated component belongs to one class in the finally obtained graph;

2. The electromagnetic signal spectrum sensing method based on the deep clustering network as claimed in claim 1, wherein the specific process of step S1 is:

s14, eliminating the dimension between the different types of parameters by using normalization, wherein the formula of the normalization is as follows:

3. The electromagnetic signal spectrum sensing method based on the deep clustering network as claimed in claim 1, wherein the specific process of the target tracking algorithm in step S3 is as follows:

and

respectively representing the speed of the target in latitude and longitude;

4. The method for sensing electromagnetic signal spectrum based on deep clustering network as claimed in claim 1,

the updating process of the set updating algorithm in step S4 is as follows: firstly, defaulting that all targets are positioned on a platform, and then segmenting the targets according to time and position information; if the distance between two targets is greater than the set value at the same moment, the two targets are not on the same platform;

the distribution process of the platform assignment algorithm in step S4 is as follows: if the distance between a new target and the targets of the known platform information is smaller than a set value at the current moment, namely the current moment is considered to be on one platform, the new target is bound to the platform to which the target nearest to the new target belongs, otherwise, a platform is newly built for storing the new target.