CN111796250A - False trace point multi-dimensional hierarchical suppression method based on risk assessment - Google Patents

Abstract

The invention relates to a multi-dimensional hierarchical suppression method of false point traces based on risk assessment, and belongs to the technical field of radar data processing. It includes four steps: extracting the multi-dimensional attribute features of the agglomerated point traces, using the unsupervised clustering algorithm and expert system to discriminate the sample set, using the artificial neural network ANN to train the point traces and their labels, and constructing the target point traces and the classifier of environmental false alarm points, calculate the score of each point trace (representing the possibility of being generated by the target), and compare the point trace score with the threshold, so as to eliminate the false point traces generated by the environmental false alarm; The method maps the multi-dimensional features of the dot trace into the dot trace quality, and completes the extraction of the target dot trace and the elimination of the clutter dot trace. The operation is simple and convenient, and the rejection rate is high. It solves the problem that the traditional point trace detection method is difficult to distinguish the target point trace and the clutter point trace, improves the detection probability of small and medium-sized targets under the condition of equal false alarm rate, and reduces the possibility of false track association and false track.

Description

A Multi-dimensional Hierarchical Suppression Method of False Traces Based on Risk Assessment

technical field

The invention belongs to the technical field of radar data processing.

Background technique

The existing active radar signal detection methods are all based on the theory that the radar signal is in stationary random white noise. By calculating the amplitude probability density distribution of radar echoes, the detection threshold under the requirement of false alarm rate is determined. Typical sea clutter amplitude is low, EP number is large, saturation is high, Doppler velocity is low, consistency is low; while target amplitude is large, EP number is small, saturation is small, Doppler velocity is high, consistency is low high degree. Although there are differences between Shanghai clutter and target echo in these dimensions, the degree of discrimination is not obvious. When the traditional constant false alarm detection method adopts a unified threshold for discrimination, it is easy to detect the heavy tail of sea clutter as the target signal output, and it is easily affected by the clutter point trace during the track initiation and track tracking process. Misstarting and mistracking occur. The patent "A Spot Trace Filtering Method Based on Support Vector Machine CN109613526A" uses spatial broadening and amplitude features as the multi-dimensional input feature parameters of the classifier; the patent "A Three-coordinate Radar Spot Trace Quality Evaluation Method CN108181620A" adopts the method of logical judgment. Point trace determination. The problems existing in the above-mentioned prior art are: the working environment of the radar is relatively complex, there is heavy trailing in the sea clutter area, and the signal variance is large, and the use of the constant false alarm method will generate a large number of false point traces, resulting in the inability of subsequent data processing. Effectively start and track radar track, generate false track, and cause track error association.

SUMMARY OF THE INVENTION

Aiming at the defects of the prior art, the purpose of the present invention is to provide a multi-dimensional hierarchical suppression method for false targets in the sea and air for risk assessment, using the extracted point trace echo features to perform quality assessment and false rejection on whether the point trace is generated by the target, It aims to improve the target detection capability in complex sea clutter environment.

To achieve the above object, the present invention adopts the following technical solutions to be realized:

(1). Collect a large amount of dot trace data in the training period;

(2) Manually discriminate point traces through track tracking, unsupervised clustering and expert systems, and perform set summation/intersection operations to distinguish target point traces and environmental false alarm point traces;

(3) Use the point trace data in the training set to train the artificial neural network, realize the multi-dimensional feature risk assessment of the point trace, and obtain the similarity between the point trace and the target point trace;

(4) According to the similarity between the point trace and the target point trace, realize the identification and elimination of false point traces.

Specifically, the above process includes a training phase and an application phase. In the training stage, multi-dimensional attribute features such as space, amplitude, frequency domain, and information domain of the point trace after the radar echo condensation are extracted, and normalized as the input of the point trace classifier; the unsupervised clustering algorithm and the expert system are used to realize For the discrimination of the sample set, a large number of point trace data are labeled, and the label is used as the output of the point trace classifier; finally, the artificial neural network ANN is used to train the point trace attributes and their corresponding labels, and the attribute characteristics of the target are learned and regularized. Mining, to establish a nonlinear mapping network between feature input and label output, so as to construct a classifier for target traces and environmental false alarm traces.

In the application stage, according to the strength of the received echo, the traditional detection method is used for target detection, the first-level judgment is completed, and the space, amplitude, frequency, and information domain features of the dot trace are extracted, and the feature extraction method and normalization method are trained. The stages are exactly the same; then the trained artificial neural network is used for the second-level judgment, the extracted high-dimensional point trace features are input, and the point trace score (the possibility of being generated by the target) is obtained to realize the quality judgment. The dot trace is used as the target to generate the dot trace, and the dot trace that does not exceed the threshold is eliminated.

The invention can adapt to the complex high sea condition environment, effectively detect the target signal, without falsely detecting the environmental echo as the target signal, thereby significantly reducing the false alarm rate. After adopting the quality evaluation and elimination of point traces based on machine learning, the number of false point traces generated by the environment is significantly reduced, while the loss of target point traces is very small. According to the statistical results, the loss rate of target point traces within the radar line of sight is less than 1 %, the dot traces generated by environmental false alarms are reduced by more than 65%, the dot trace quality (correct rate) is improved; the track calculation time is reduced by 23%, and the false tracks are reduced by more than 50%, which proves the effectiveness of the present invention.

Description of drawings

Figure 1. Flow chart of multi-dimensional hierarchical suppression of false targets based on risk assessment.

Figure 2. Input and output of multi-dimensional hierarchical suppression of false targets based on risk assessment.

Figure 3. The target traces extracted after the multi-dimensional hierarchical suppression of false targets.

Detailed ways

The present invention will be further explained below in conjunction with the accompanying drawings.

The input information of the dot trace used includes the spatial information of the dot trace, such as distance, azimuth, elevation angle, the number of point traces crossing the threshold (EP number) during condensation, saturation (the ratio of the EP number to the total number of distance units in the sector), Range broadening, azimuth broadening; amplitude information such as the range unit amplitude, CFAR estimated background amplitude, and signal-to-noise ratio; frequency domain information such as the main channel number, the proportion of the agglomeration point trace located in the main channel number, the number of channels that pass the threshold, the number of channels that pass the threshold Channel sequence number variance, intra-channel signal consistency; signal processing discrimination information such as regional attributes (sea clutter area, ground object area, noise area), whether it is in the sidelobe area of the tracking target, a total of 21 dimensions. The relevant parameters are defined as follows:

Distance refers to the distance between the current point trace and the radar

Bearing refers to the bearing of the current trace in the radar coordinate system

Elevation angle refers to the elevation angle of the current trace in the radar coordinate system

Amplitude refers to the average amplitude of the current trace

Background amplitude refers to the amplitude of the background where the current trace is located

SNR refers to the SNR of the current dot trace SCNR﹦Amp－BackAmp;

The number of over-threshold points refers to the number of over-threshold points that form the current trace, and record OutNum﹦N in the process of trace condensation;

Saturation refers to the proportion of the threshold crossing points forming the current trace to the trace frame

SatDeg=N/((DisWid/DisSample)·(AziWid/AziSample));

The distance extension refers to the distance extension of the current point trace DisWid﹦SDis－EDis;

The starting distance SDis refers to the starting distance of the current track, and the ending distance EDis refers to the ending distance of the current track.

Azimuth widening refers to the azimuth span of the current trace AziWid﹦SAzi－EAzi;

The starting azimuth SAzi refers to the starting azimuth of the current trace, and the ending azimuth EAzi refers to the ending azimuth of the current trace.

The area attribute refers to what type of area the current point trace is located in, including the ground object area, the sea clutter area, the cloud and rain clutter area, the noise area, and the over-the-horizon area. The main channel number refers to the main component channel of the current point trace. Assuming that N over-threshold points belong to m output channels respectively, the channel numbers are Chn ₁ ,...Chn _m , and the number of over-threshold points of each channel are OutNums ₁ ,...OutNums _m , then the main channel numbers are:

MainChn ₁ ﹦Chn _x , OutNums _x ﹦max(OutNums ₁ ,…OutNums _m ),

The main channel ratio refers to the ratio of the threshold crossing points of the main component channels MainRate﹦OutNums _x /N,

The number of channels refers to the number of channels that make up the current trace ChnNums﹦m,

The channel consistency refers to the standard deviation ArgDeg﹦std(Chn ₁ ,…Chn _m ) of the channel numbers that make up the current trace,

The side lobe identification refers to whether the current point trace may be a point trace formed by a side lobe. When the dot trace is located in the side lobe area of a certain dot trace and satisfies certain criteria, the dot trace is considered as a side lobe dot trace. The judgment criteria are as follows:

(1) The point trace needs to be located in the side lobe area of a certain point trace (main lobe point trace); the side lobe area refers to the position of the main lobe point trace as the center, and the distance to the side lobe refers to the same orientation as the target, and the distance increases or decreases. The corresponding position, the azimuth sidelobe refers to the position at the same distance from the target, and the azimuth increase or decrease corresponds to the position;

(2) The amplitudes of the dot trace and the main lobe dot trace need to satisfy the relationship between the primary and secondary ratios;

(3) The dot trace and the main lobe dot trace need to be located in the similar output channel.

There are 21 input parameters in total.

The specific process flow of the present invention is as shown in Figure 1, and its implementation steps are as follows: Step 1, adopt constant false alarm detection and dot trace condensation to form dot traces and extract dot trace features;

Step 2. Correlate the point track data to form a track, and select a track with an association longer than 20 laps as the target track;

Step 3. Find the cluster center and the category number of the point trace through the Kmeans unsupervised clustering algorithm;

Step 4. Send the clustered point trace cluster into the expert system, and manually distinguish the target point trace and the clutter trace;

Step 5, the point traces of step 2 and step 4 are obtained by the union of the point traces to obtain the target traces, and the summation is obtained to obtain the clutter traces;

Step 6. Divide the dot trace feature and its category (including label data) in step 5, randomly select 80% of the samples as the training set, and the remaining samples as the verification set;

Step 7. Send the sample set to the artificial neural network for training, and use the verification set to verify the correctness;

Step 8, take the actual point trace as input, and use the artificial neural network trained in step 7 to calculate the similarity with the target point trace;

Step 9: Set a threshold for the number of dot traces, filter out the dot trace data with low similarity to the target, and output the target dot trace.

Wherein, the specific steps of step 3 are:

Step 3.1: Set the number of categories to N, N refers to the sum of the estimated target category and the false categories generated by the environment. For example, if there are 80 targets, it means that there are 80 categories of targets estimated, and the number of false categories is 20. At this time, N is selected as 100. In the evaluation of dot trace quality, 80 categories of targets are given a high score and 20 categories of false are given a low score. After N is given, randomly select N sample points in the dot trace as the category center of these N types; where X is the category attribute, which represents the central value of the attribute feature of this type of dot trace, and the number of dot traces is L.

X _k = {x ₁ , x ₂ ... x _N }, i=1, 2...L, k=1, 2...N

Step 3.2: Calculate the Euclidean distance between all point traces and all categories, and use the distance to determine the category of the point trace; where P is the feature dimension of the point trace sample.

T _k =arg min _k {|x _i -X _k |}

Step 3.3: Traverse all categories, and recalculate the center of the category using all traces belonging to each category; where N _k represents the number of traces closest to the center of the kth category;

Step 3.4: Repeat steps 3.2-3.3 cyclically, and perform Niter times in total;

Step 3.5: The Euclidean distance is used to calculate the distance between the categories, and the two categories whose category distance is less than the threshold are merged into a new category; if the distance between the mth class and the nth class is less than the threshold, according to the number of samples to which they belong. The numbers are weighted, the center of the s-th class is calculated, and the m-th and n-th class centers are removed.

Step 3.6: Calculate the variance of each feature attribute in the category. If the variance is greater than the threshold, the category will be split into two categories, and the centers of the new two categories will copy the center of the original category, and the point trace attribute dimension with the largest variance (No. h attributes), plus or minus the standard deviation of the trace attributes;

Step 3.7: When the merging of classes no longer occurs in step 5, and the splitting of categories no longer occurs in step 6, the entire unsupervised clustering process ends, otherwise steps 4-6 are performed;

The specific steps of step 7 are:

Step 7.1: Read the normalized input feature x of the data and the label v saved by the expert system;

Step 7.2: Set the neural network structure and activation function type, and initialize the network weights randomly; set the initialization coefficients of the three-layer network to be ω _i , i=1, 2, and 3, respectively, and the excitation thresholds to be bi , _i =1, 2 , 3;

Step 7.3: Set the network training step size h, the regularization coefficient a, the momentum coefficient M, and the maximum training times Niter;

Step 7.4: Network forward calculation to obtain prediction results

g(x)=x

The LeakyReLu function is:

Step 7.5: Calculate the difference between the prediction result and the true value of the label, called the error δ, define the square cost function as the mean of the square errors of the M samples, and the coefficient 0.5 is to offset the factor generated during the derivation;

Step 7.6: Calculate the network backward, calculate the update coefficient of each node through the cost function and the chain rule of derivation;

The derivative of the LeakyReLu function is:

Step 7.7: Calculate the momentum of the update coefficients of the network nodes;

dω _i =Mdω _i +Δω _i

db _i =Mdb _i + _Δbi , M<1

Step 7.8: Perform micro-forgetting on the weight value of the network node, and update the weight according to the result of step 7;

ω _i =(1-α)ω _i +dω _i

b _i =(1-α)b _i +db _i , α<<1

Step 7.9: Repeat steps 7.4-7.8 until the network error is less than the threshold or the maximum number of training Niter is reached.

Claims (3)

1. A false point trace multi-dimensional hierarchical suppression method based on risk assessment is characterized by comprising the following steps:

(1) collecting a large amount of trace point data in a training period;

(2) manually distinguishing the point traces by track tracking, unsupervised clustering and an expert system, and carrying out aggregate merging/intersection calculation to distinguish a target point trace and an environmental false alarm point trace;

(3) training the artificial neural network by using trace point data in the training set, realizing multi-dimensional characteristic risk assessment of trace points, and obtaining the similarity between the trace points and a target trace point;

(4) and according to the similarity between the trace points and the target trace points, realizing the identification and elimination of false trace points.

2. The multi-dimensional hierarchical suppression method for false traces based on risk assessment as claimed in claim 1, wherein: the unsupervised clustering is as follows: and judging the distance between the class centers in the iterative process, combining the class centers with the short distance into 1 class, and splitting the class with the large variance into 2 new classes.

3. The multi-dimensional hierarchical suppression method for false traces based on risk assessment as claimed in claim 1 or claim 2, wherein: the artificial neural network training adopts a regularization and momentum training method in the training process.

Images