Method of Improving Radar Signal Detection Rate Using Window Function

Abstract

The present invention provides a method for improving the detection rate of radar signals by utilizing a window function. By multiplying the window function during the convolution process, not only can the sidelobe energy be suppressed, but also the problem of the inseparability of target signals and noise signals is solved. On this basis, a real part detection model and a corresponding neural network model are established, the neural network model is trained, and the trained neural network model is used to detect echo signals to distinguish target signals and noise signals in the echo signals. Due to the use of real part detection and deep learning methods, the signal-to-noise ratio is further improved, thereby improving the detection rate of radar signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal detection, and in particular relates to a method for improving the detection rate of radar signals by utilizing a window function. Background Art

[0002] The working process of traditional pulse Doppler radar is generally as follows: the radar transmitter sends a specifically modulated pulse signal through the antenna. The signal propagates through the air and returns to the radar after encountering a target, and is received by the receiver antenna. Among them, the pulse signal emitted by the radar is mainly divided into two parts: one is the baseband signal, also known as the modulation signal, which is a sinusoidal wave signal with a specific waveform formed by modulation such as amplitude modulation, phase modulation, or frequency modulation; the other is the carrier signal, which has the function of carrying the modulation signal. Because the modulation signal has a very low frequency (usually within 1GHz) and a very wide bandwidth, it is easy to interfere with each other, which is not conducive to long-distance transmission. However, by transporting it with a higher frequency carrier signal (usually around 10GHz), the above-mentioned shortcomings of the modulation signal can be effectively solved.

[0003] Due to energy losses from signal propagation and target reflection, the signals received by the receiver are often very weak. Therefore, they must be amplified by a low-noise amplifier before subsequent demodulation, detection, and display can be performed. However, due to the presence of resistor thermal noise in the amplifier circuit, noise is inevitably added to the signal during the amplification process, affecting the detection of weak signals.

[0004] The basis for radar to measure the distance, speed and other aspects of a target is whether the radar can correctly judge whether it has received a signal, that is, whether the signal can be detected correctly. Usually, the Neyman-Pearson criterion is used, that is, at a given false alarm rate P fa Under the conditions of SNR and signal-to-noise ratio, the detection rate P d Reach the maximum.

[0005] False alarm rate P fa It refers to the ratio of noise to signal during signal detection. Since noise is generated continuously, even if the false alarm rate reaches 10 -6 , that is, if there is only one false alarm among 1 million detection points, then the alarm will also give a false alarm once every ten seconds. Therefore, in engineering applications, the false alarm rate should be as low as possible. d It refers to the proportion of correctly detected signals to all signals. It is the guarantee for whether all signals can be recognized, and it is also the key to whether the target can be detected correctly and timely.

[0006] Among existing radar signal detection methods, the most commonly used method is the modulus length detection method. The detection steps are roughly as follows:

[0007] (1) Since the echo mode length follows the Rice distribution, the false alarm rate is calculated using the hypothesis testing method;

[0008] (2) Obtain the signal size and signal convolution value corresponding to the acceptable false alarm rate from the probability density function;

[0009] (3) After receiving the echo signal, pass the echo signal through a matched filter to obtain a convolution value;

[0010] (4) Perform energy detection, i.e., detect the size of the convolution value;

[0011] (5) By setting a threshold, the echo points above the threshold are intercepted and considered as signals.

[0012] Using existing detection methods, the false alarm rate can reach 10 -6 , the detection rate can reach 83%. However, due to the limitations of the method itself, the detection rate is difficult to be further improved. Summary of the Invention

[0013] The present invention is made to solve the above problems and aims to provide a method for improving the detection rate of radar signals by using window functions and deep learning. The present invention adopts the following technical solutions:

[0014] The present invention provides a method for improving the detection rate of radar signals by using a window function, which is characterized by comprising:

[0015] Step S1, obtaining an echo signal to be detected, the echo signal including a plurality of echo points;

[0016] Step S2, convolving the echo signal and multiplying it by a window function during the convolution process to obtain a convolution value of each echo point;

[0017] Step S3, establishing a real part detection model;

[0018] Step S4, establishing a corresponding neural network model based on the real part detection model;

[0019] Step S5, training the neural network model;

[0020] Step S6: Use the trained neural network model to detect the echo signal and distinguish the target signal and the noise signal in the echo signal.

[0021] Wherein, step S3 includes the following sub-steps:

[0022] Step S3-1: Model the echo point as including m signal points, where each signal point is a complex signal including a real part and an imaginary part:

[0023] z=ξ+jζ,

[0024] Where ξ and ζ are independent and identically distributed in the normal distribution N(0,σ 2 ), j is the imaginary unit;

[0025] Step S3-2, splitting the echo point whose convolution value reaches a predetermined threshold into m signal points;

[0026] Step S3-3: Take the m signal points as samples and establish a hypothesis test:

[0027] H0:x~N n (0,Σ),

[0028] H1:x~N n (μ, Σ),

[0029] Wherein, μ is the square of the modulus of the signal point, and Σ is the covariance matrix of the product of the target signal and the convolution function.

[0030] The method for improving the radar signal detection rate by using a window function provided by the present invention may also have such a technical feature, wherein the window function is a Hanning window:

[0031]

[0032] The method for improving the detection rate of radar signals by using a window function provided by the present invention may also have such a technical feature, wherein, in step S2, convolution is performed using the following formula:

[0033]

[0034] Where x 2 +y 2 =1, k is the frequency modulation slope.

[0035] The method for improving the radar signal detection rate by using a window function provided by the present invention may also have the following technical feature: m=1201.

[0036] The method for improving the radar signal detection rate by using a window function provided by the present invention may also have such a technical feature, wherein, in step S5, a plurality of signal points that are false alarms after adding the window function are formed into an image according to their position coordinates, and labels are added as training data to train the neural network model.

[0037] Functions and effects of the invention

[0038] According to the method of improving the detection rate of radar signals by using a window function of the present invention, a window function is multiplied during the convolution process, thereby not only suppressing the sidelobe energy, but also solving the problem that the target signal and the noise signal are inseparable. On this basis, a real part detection model and a corresponding neural network model are established, the neural network model is trained, and the trained neural network model is used to detect the echo signal to distinguish the target signal and the noise signal in the echo signal. Due to the use of real part detection and deep learning methods, the signal-to-noise ratio is further improved, thereby improving the detection rate of radar signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of a method for improving radar signal detection rate by using a window function in an embodiment of the present invention;

[0040] Figure 2 is a flow chart of step S3 in an embodiment of the present invention;

[0041] Figure 3 1 is a graph showing the changing trend of the accuracy of the training set using different window functions in an embodiment of the present invention;

[0042] Figure 4 1 is a graph showing the changing trend of the cost function using different window functions on the training set in an embodiment of the present invention;

[0043] Figure 5 1 is a graph showing the changing trend of the accuracy of different window functions on the validation set in an embodiment of the present invention;

[0044] Figure 6 is a graph showing the changing trend of the cost function using different window functions on the validation set in an embodiment of the present invention;

[0045] Figure 7 is a graph showing a change trend of the accuracy of the Hanning window at different thresholds according to an embodiment of the present invention;

[0046] Figure 8 is a graph showing a change trend of the cost function of the Hanning window at different thresholds in an embodiment of the present invention;

[0047] Figure 9 1 is a schematic diagram of a sequence of real and imaginary parts of an echo signal in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the following is a detailed description of the method of using a window function to improve the detection rate of radar signals in accordance with the present invention, with reference to the embodiments and accompanying drawings.

[0049] <Example>

[0050] This embodiment provides a method for improving the detection rate of radar signals by using a window function, which is used to distinguish target signals from noise signals in false alarm echo signals, thereby improving the detection rate.

[0051] Figure 1 4 is a flow chart of a method for improving the radar signal detection rate by using a window function in this embodiment.

[0052] like Figure 1 As shown in FIG, the method for improving the detection rate of radar signals by using a window function specifically includes the following steps:

[0053] Step S1: Acquire an echo signal to be detected, where the echo signal includes a number of echo points.

[0054] Due to the confidential nature of military radar detection data and the extremely small number of target signals in practice, it is difficult to directly obtain a large amount of real data. However, after studying the distribution of radar signals, it was found that the noise in radar signals is mostly Gaussian white noise, while the target signals are mostly linear frequency modulation signals. Therefore, in this embodiment, based on this research result, a large amount of simulated data for testing was generated.

[0055] Figure 9 is a sequence diagram of the real and imaginary parts of the echo signal in this embodiment.

[0056] like Figure 9 As shown in the figure, the real and imaginary parts of the noise are both normally distributed with a mean of 0 and a variance of 0.5, while the target signal is a variable with an initial phase that only changes with time. For the sake of generality, the initial phase of the target signal is set to 0. At this time, the echo signal with the target signal passing through the matched filter is equivalent to a complex signal with a normal distribution with a non-zero mean as the real part and a normal distribution with a zero mean as the imaginary part.

[0057] Step S2: Convolve the echo signal and multiply it by a window function during the convolution process to obtain the convolution value of each echo point.

[0058] According to the above research results, the target signal is mainly a linear frequency modulation signal, that is, within a pulse cycle, the frequency of the signal changes with time, and its complex signal expression is:

[0059]

[0060] In the formula, K = B / T, that is, the bandwidth of the signal divided by the time width (abbreviated as time width), which is called the frequency modulation slope.

[0061] The noise signal is mainly a Gaussian white noise signal. “Gaussian” means that its statistical characteristics generally conform to the Gaussian distribution, and “white” means that its power spectrum density is constant. Its complex expression is:

[0062] n(t)=A+jB,

[0063] In the formula, A and B both have a mean of 0 and a variance of σ 2 The standard normal distribution of .

[0064] The echo signal includes the target signal and the noise signal.

[0065] The convolution is performed using the following convolution function:

[0066]

[0067] Where x 2 +y 2 =1, k is the frequency modulation slope.

[0068] In this embodiment, the following window functions are used for horizontal comparison:

[0069] 1. Hanning Window:

[0070]

[0071] 2. Non-continuous window function:

[0072]

[0073] 3. Continuous window function:

[0074]

[0075] 4. Slash window function:

[0076]

[0077] 5. Window function with all values ​​set to 1 (equivalent to no windowing):

[0078] ω(n)=1。

[0079] Step S3: establishing a real part detection model.

[0080] Figure 2 This is a flow chart of step S3 in this embodiment.

[0081] like Figure 2 As shown, step S3 specifically includes the following sub-steps:

[0082] Step S3-1: Model the echo point as including m signal points, where each signal point is a complex signal including a real part and an imaginary part:

[0083] z=ξ+jζ,

[0084] Where ξ and ζ are independent and identically distributed in the normal distribution N(0,σ 2 ), j is the imaginary unit.

[0085] In this embodiment, the echo point whose convolution value reaches the predetermined threshold contains 1201 variables z, which can be proved that at each moment, the variable z and the convolution function After multiplication, it is still an independent normal distribution, that is:

[0086]

[0087] Obviously, (xξ+yζ), (yξ+xζ) in the formula are independent and identically distributed in the normal distribution N(0,σ 2 ), so without loss of generality, for the sake of symbolic simplicity, the variable after the product is still recorded as z.

[0088] Step S3-2: split the echo points whose convolution values ​​reach a predetermined threshold into the original m signal points.

[0089] Step S3-3, taking m signal points as samples, establish hypothesis test:

[0090] H0:x~N n (0,Σ),

[0091] H1:x~N n (μ, Σ),

[0092] In the formula, μ is the product of the signal s(t) and the convolution function, that is, the square of the modulus of the signal point, and Σ is...

[0093] Without adding the window function in step S2, the above method does not work. The following is a simple demonstration:

[0094] Assume x i (i=1,2,…,n) are independent and identically distributed in N(μ,σ 2 ), let y = ∑x i ~N(nμ,nσ 2 ), obviously y represents the convolution value of the target signal point.

[0095] x i (i=1,2,…,n-1) and y form a multivariate normal distribution, whose mean vector and covariance matrix are as follows:

[0096]

[0097] make:

[0098]

[0099] The conditional distribution X|y~N(μ 11·2 ,Σ 11·2 ):

[0100]

[0101]

[0102] From the above formula, we can see that no matter whether the signal is 0 or not, it cannot be distinguished through hypothesis testing.

[0103] In step S2, a window function is multiplied during the convolution process. On the one hand, it can suppress the sidelobe energy, and on the other hand, it solves the problem that the target signal and the noise signal are inseparable:

[0104] X * =KX T

[0105] Where, k i (i=1, 2, ... n) is a constant depending on the window function.

[0106] That is x i i=1,2,…,n are independently distributed in make:

[0107]

[0108] The conditional distribution can also be obtained

[0109]

[0110] make:

[0111] M=∑k i ,

[0112] but:

[0113]

[0114] Where y is the convolution value of the echo, the following formula is obviously true:

[0115] y=y s +y n

[0116] Where y s ,y n Represent the noise convolution value and signal convolution value respectively, then:

[0117] y s =Mμ

[0118] Then we can get:

[0119]

[0120] As can be seen from the above formula, under the condition of known echo convolution value, the mean of the echo distribution is completely determined by the target signal and the noise signal. More importantly, the mean of the echo distribution is related to the target signal. Therefore, various methods can be used for detection and to distinguish the target signal from the noise signal.

[0121] In addition, both the above-mentioned discontinuous window function and the continuous window function can theoretically make M=0, so that the mean information of the echo distribution can be retained to the greatest extent.

[0122] Through the above steps, a real part detection model for echo signal detection is established.

[0123] Step S4: establishing a corresponding neural network model based on the real part detection model.

[0124] Step S5: training the neural network model.

[0125] In this example, a 154-layer deep convolutional neural network was used. Multiple signal points that were false alarms after windowing were mapped to a 400*400 image based on their positions. These points were manually labeled and used as training data before being input into the neural network model for training. The different windowing functions were also used to facilitate horizontal comparison.

[0126] Step S6: Use the trained neural network model to detect the echo signal and distinguish the target signal and the noise signal in the echo signal.

[0127] Figure 3 This is a trend chart of the accuracy of the training set using different window functions in this embodiment.

[0128] Figure 4 3 is a graph showing the changing trend of the cost function using different window functions on the training set in the embodiment.

[0129] Figure 5 This is a trend chart of the accuracy of different window functions on the validation set in this embodiment.

[0130] Figure 6 3 is a graph showing the changing trend of the cost function using different window functions on the validation set in this embodiment.

[0131] like Figure 3-6 As shown, in this embodiment, a total of 2,000 training times are performed, and the accuracy of the models using different window functions and the changing trends of the cost functions with the number of training times are recorded. It can be seen that the Hanning window has a better effect. Therefore, the Hanning window with better performance is further compared vertically in terms of threshold size.

[0132] Figure 73 is a graph showing the changing trend of the accuracy of the Hanning window at different thresholds in this embodiment.

[0133] Figure 8 3 is a graph showing the changing trend of the cost function of the Hanning window at different thresholds in this embodiment.

[0134] like Figure 7 and Figure 8 As shown in Figure 1, after adding the Hanning window, the thresholds are set to 10, 20, 30, 40, and 50 respectively, and the changing trends of the accuracy and cost function are recorded. The results are also shown in Table 1.

[0135] Table 1 Effects of different window functions at different thresholds

[0136]

[0137] As can be seen from Table 1, the best result can reach 94% in the existing threshold detection standard. That is to say, if the same 13dB signal is detected, the method of this embodiment can reach 10 -7 The false alarm rate is ten times higher than that of existing methods, and the detection rate is increased to 92%, which is significantly better than the 83% of existing detection methods.

[0138] Example Function and Effect

[0139] According to the method for improving the radar signal detection rate by using a window function provided in this embodiment, the window function is multiplied during the convolution process, thereby not only suppressing the sidelobe energy, but also solving the problem of the inseparability of the target signal and the noise signal. On this basis, a real part detection model and a corresponding neural network model are established, the neural network model is trained, and the trained neural network model is used to detect the echo signal to distinguish the target signal and the noise signal in the echo signal. Due to the use of real part detection and deep learning methods, the signal-to-noise ratio is further improved, thereby improving the detection rate.

[0140] In the embodiment, through comparative analysis, it is found that the best detection effect is the Hanning window. By using the method of this embodiment and the Hanning window, the best result can reach 94% in the existing threshold detection standard. That is to say, if the same 13dB signal is detected, the method of this embodiment can reach 10 -7 The false alarm rate is improved tenfold compared with the existing method, and the detection rate is increased to 92%, which is significantly better than the 83% of the existing detection method. Therefore, the method of this embodiment significantly improves the detection rate of radar signals.

[0141] The above embodiments are only used to illustrate specific implementations of the present invention, and the present invention is not limited to the description scope of the above embodiments.

Claims

1. A method for improving the detection rate of radar signals using a window function, characterized in that: include: Step S1, obtaining an echo signal to be detected, the echo signal including a plurality of echo points; Step S2, convolving the echo signal and multiplying it by a window function during the convolution process to obtain a convolution value of each echo point; Step S3, establishing a real part detection model; Step S4, establishing a corresponding neural network model based on the real part detection model; Step S5, training the neural network model; Step S6: Use the trained neural network model to detect the echo signal and distinguish the target signal and the noise signal in the echo signal. Wherein, step S3 includes the following sub-steps: Step S3-1: Model the echo point as including m signal points, where each signal point is a complex signal including a real part and an imaginary part: z=ξ+jζ, Where ξ and ζ are independent and identically distributed in the normal distribution N(0,σ 2 ), j is the imaginary unit; Step S3-2, splitting the echo point whose convolution value reaches a predetermined threshold into m signal points; Step S3-3: Take the m signal points as samples and establish a hypothesis test: H0:x~N n (0,Σ), H1:x~N n (m,S), Wherein, μ is the square of the modulus of the signal point, Σ is the covariance matrix of the product of the target signal and the convolution function, In step S2, the window function is multiplied during the convolution process to suppress the sidelobe energy on the one hand and solve the problem that the target signal and the noise signal are inseparable on the other hand: Let X * =KX T , Where, k i (i=1,2,…n) is a constant that depends on the window function, where x i (i=1,2,…,n) are independently distributed in make: The conditional distribution can also be obtained make: but: Where y is the convolution value of the echo, the following equation holds: y=y s +y n Where y s ,y n Represent the noise convolution value and signal convolution value respectively, then: y s =Mμ Then we can get: It can be seen that under the condition of known echo convolution value, the mean of echo distribution is completely determined by the target signal and noise signal. The mean of echo distribution is related to the target signal. In step S5, a plurality of signal points that are false alarms after adding the window function are formed into a picture according to their position coordinates, and are labeled as training data, and the training data are input into the neural network model for training.

2. The method for improving radar signal detection rate by using a window function according to claim 1, wherein: in, The window function is a Hanning window:

3. The method for improving radar signal detection rate by using a window function according to claim 1, wherein: in, In step S2, convolution is performed using the following convolution function: Where x 2 +y 2 =1, k is the frequency modulation slope.

4. The method for improving radar signal detection rate by using a window function according to claim 1, wherein: in, m＝1201。

Images