A new adaptive sonar target detection method and computer device based on neural network

Abstract

The present invention discloses a novel adaptive sonar target detection method and computer equipment based on a neural network, which belongs to the field of underwater acoustic detection technology, and solves the problem that it is difficult to achieve a relatively robust detection performance under a relatively complex background environment using a CFAR detection method for detection. The method of the present invention comprises: performing matched filtering processing on the received sonar target signal; performing detection processing on the filtered sonar target signal to obtain a detection unit and a reference unit; classifying the sonar target signal according to the reference unit and using a neural network to obtain the background clutter type of the detection unit; selecting the best detection method according to the background clutter type of the detection unit; and using the best detection method to complete the detection of the sonar target. The present invention is suitable for the detection of sonar targets in underwater acoustic detection.

Description

Technical Field

[0001] The present application relates to the field of underwater acoustic detection technology, and in particular to a novel adaptive sonar target detection method based on a neural network. Background Art

[0002] Constant false alarm rate detection (CFAR) is an effective method in sonar target detection. In this method, the false alarm rate will be used as a priori parameter and remain unchanged. The classic and commonly used CFAR methods mainly include cell average constant false alarm detector (CellAveraging CFAR, CA-CFAR), ordered statistics CFAR detector (Ordered Statistics CFAR, OS-CFAR), automatic screening average constant false alarm detector (Auto Censored Mean Level Detector CFAR, ACMLD-CFAR), the Greatest Of CFAR (GO-CFAR), the Smallest Of CFAR (SO-CFAR), etc. These commonly used constant false alarm detectors have high performance only in one or more special occasions. For example, in a uniform clutter background, the detection performance of CA-CFAR is better than other detectors, but the performance in multi-target and clutter edge will drop sharply; in a multi-target environment, OS-CFAR and ACMLD-CFAR have better detection capabilities, among which the ACMLD method is derived on the basis of the OS method. The ACMLD method can appropriately delete the strong discrete reverberation reference units that cause non-uniformity and make full use of more reference unit data. Therefore, the detection performance will not show a sharp decline with the increase of the number of interfering targets; in a low-power environment at the edge of clutter, the false alarm control capability of SO-CFAR is better than other detectors, and the false alarm rate is also lower than other detectors. In a high-power environment at the edge of clutter, the false alarm control capability of GO-CFAR is better than other detectors, and the false alarm rate is also lower than other detectors.

[0003] Artificial neural network is a highly simplified biological neural network. It is formed by interconnecting many computing units. It has a strong nonlinear fitting function and can achieve a certain degree of self-learning ability and extraordinary memory ability through computers. There are many artificial neural network models now, including radial basis neural network, convolutional neural network, recurrent neural network, single-layer perceptron, multi-layer perceptron, BP neural network are all typical models. Neural network has been effectively applied in many engineering fields, including underwater acoustic signal processing, for example: applying neural network in spectrum estimation, adaptive filtering direction, sonar array signal processing and signal reconstruction, sonar target detection based on neural network, etc. At present, the CFAR method based on neural network algorithm is still widely studied.

[0004] In the prior art, such as Figure 1 As shown, it is difficult to achieve a relatively robust detection performance in a relatively complex background environment by using a CFAR detection method for detection. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that it is difficult to obtain relatively robust detection performance under relatively complex background environments by using a CFAR detection method in the prior art, and to provide a novel adaptive sonar target detection method and computer device based on a neural network.

[0006] The present invention is implemented by the following technical solutions. On the one hand, the present invention provides a novel adaptive sonar target detection method based on a neural network, the method comprising:

[0007] Perform matched filtering on the received sonar target signal;

[0008] Performing detection processing on the filtered sonar target signal to obtain a detection unit and a reference unit;

[0009] According to the reference unit, the sonar target signal is classified by using a neural network to obtain the background clutter type of the detection unit;

[0010] Selecting the best detection method according to the background clutter type of the detection unit;

[0011] The optimal detection method is used to complete the detection of sonar targets.

[0012] Furthermore, the detection processing method includes: square-law detection or linear detection.

[0013] Furthermore, the acquisition method of the detection unit specifically includes:

[0014] According to the filtered sonar target signal, obtaining an in-phase component and an orthogonal component of the sonar target signal;

[0015] The detection unit is acquired according to the in-phase component and the orthogonal component.

[0016] Furthermore, the classification specifically includes: dividing the sonar target signal into four types: a uniform background environment, a high-power area at the edge of clutter, a low-power area at the edge of clutter, and a multi-target background environment.

[0017] Furthermore, the sonar target signal is classified according to the reference unit and using a neural network, specifically: the reference unit is input into a clutter background recognition and classification network.

[0018] Furthermore, the clutter background recognition and classification network adopts the Levenberg-Marquardt algorithm for feedback training.

[0019] Furthermore, the selecting the best detection method according to the background clutter type of the detection unit specifically includes:

[0020] If the background clutter type is a uniform background environment, select the CA-CFAR detector;

[0021] If the background clutter type is a high power area at the edge of the clutter environment, select the GO-CFAR detector;

[0022] If the background clutter type is a low-power area at the edge of the clutter, a SO-CFAR detector is selected;

[0023] If the background clutter type is a multi-target background environment, the ACMLD-CFAR detector is selected.

[0024] On the other hand, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the steps of the novel adaptive sonar target detection method based on neural network as described above are executed.

[0025] The beneficial effects of the present invention are as follows: a new adaptive sonar target detection method based on a neural network is adopted, the sonar target signal is classified and identified using a neural network, and the optimal CFAR detection method is adaptively selected, which can not only improve the anti-interference ability of the system, but also achieve the purpose of accurate detection.

[0026] The invention is suitable for detecting sonar targets in underwater acoustic detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0028] Figure 1 It is a practical operation flow chart of prior art sonar target detection;

[0029] Figure 2 It is the structural block diagram of CFAR detector;

[0030] Figure 3 It is the BP neural network feedback mechanism;

[0031] Figure 4 It is a structural diagram of the novel sonar signal detector of the present invention. DETAILED DESCRIPTION

[0032] Implementation method 1: Figure 1-4 As shown, a novel adaptive sonar target detection method based on a neural network comprises:

[0033] Perform matched filtering on the received sonar target signal;

[0034] Performing detection processing on the filtered sonar target signal to obtain a detection unit and a reference unit;

[0035] According to the reference unit, the sonar target signal is classified by using a neural network to obtain the background clutter type of the detection unit;

[0036] Selecting the best detection method according to the background clutter type of the detection unit;

[0037] The optimal detection method is used to complete the detection of sonar targets.

[0038] In this implementation, a neural network is combined with a sonar constant false alarm detection technology, and the most appropriate constant false alarm detection method is adaptively selected using the neural network training results.

[0039] After measuring the original signal, the received sonar target signal will first be processed by matched filtering in order to improve the signal-to-noise ratio; after the detection unit is obtained by square law detection, it is input into the clutter background recognition and classification network, and the best detection method is selected according to the classification result to complete the accurate detection of the sonar target.

[0040] Embodiment 2: This embodiment further limits the novel adaptive sonar target detection method based on neural network described in embodiment 1. In this embodiment, the detection processing method is further limited, specifically including: square law detection or linear detection.

[0041] In this implementation, the filtered sonar target signal is subjected to detection processing, and both detection methods can obtain the detection unit of the constant false alarm detection system.

[0042] Embodiment 3: This embodiment further limits the novel adaptive sonar target detection method based on neural network described in embodiment 1. In this embodiment, the acquisition method of the detection unit is further limited, specifically including:

[0043] According to the filtered sonar target signal, obtaining an in-phase component and an orthogonal component of the sonar target signal;

[0044] The detection unit is acquired according to the in-phase component and the orthogonal component.

[0045] In this embodiment, after the filtered sonar target signal is detected, the in-phase component I(t) and the quadrature component Q(t) of the signal can be analyzed, and D(t) is the output result after I(t) and Q(t) are detected. Common detection methods include square law detection, which is expressed as:

[0046] D(t)=I 2 (t)+Q 2 (t) (1)

[0047] Linear detection, the expression is:

[0048]

[0049] Figure 2 D is the detection unit of the constant false alarm detection system. The detection threshold of the system is determined by the reference units x on both sides. i (i=1,2…,n) and y i (i=1,2…,n) is related to the multiplication factor T.

[0050] It should be noted that the purpose of the detection unit is to estimate the background noise power level. The specific calculation method is related to the form of the selected CFAR algorithm. After Z is obtained, the detection threshold of the system is obtained by the formula S=Z*T. Then this value is compared with the detection unit to determine whether there is a target. Corresponding distance units should be set on both sides of the detection unit, generally called protection units, to prevent the target signal from entering the reference units on both sides, increase the noise power, and thus increase the detection threshold of the system. When the threshold is increased, the threshold will increase, and the target will not be detected after it is greater than the target power.

[0051] Embodiment 4: This embodiment further limits the novel adaptive sonar target detection method based on a neural network described in Embodiment 1. In this embodiment, the classification is further limited, specifically including:

[0052] The sonar target signals are divided into four types: uniform background environment, high power area at the edge of clutter environment, low power area at the edge of clutter environment and multi-target background environment.

[0053] It should be noted that under the condition of uniform noise distribution, according to the characteristics of independent and identically distributed sampling of reference units, the false alarm probability expression can be obtained as follows:

[0054]

[0055] In formula (3), μ is the exponential distribution parameter, and the random variable Z can be obtained from the signal in the reference window. Z (z) is the probability density function of Z, M Z (u) is the moment generating function of Z. When u = T / (1+SNR)μ, the detection probability p of the system in a uniform noise background can be obtained. d ,Right now:

[0056] p d =M Z (u)| u＝T(1+SNR)μ (4)

[0057] In the formula, SNR is the signal-to-noise ratio. From the above analysis, it can be seen that the false alarm probability p of constant false alarm detection can be solved by calculating the moment generating function of the background noise estimation. f and the detection probability p d .

[0058] First, the principle of CA-CFAR is analyzed. The detector uses the values ​​of the reference cells in the front and rear reference windows N = 2n to calculate the mean of its samples and use it as the total background noise power, that is:

[0059]

[0060] Where Z CA represents the total background noise power estimate.

[0061] The probability density function PDF of the Γ distribution is:

[0062] f(x)=β -α x α-1 e -xβ / Γ(α) x≥0,α≥0,β≥0 (6)

[0063] Γ(α)=(α-1)! (7)

[0064] If a random variable X follows a Γ distribution, it can be expressed as the following expression X~G(α,β), and its moment generating function is as follows:

[0065] M X (u) = (1 + βu) -α (8)

[0066] The exponential distribution is a special case of the Γ distribution with α = 1. In a uniform noise environment, we know that x i ~G(1,μ),y i ~G(1,μ) and Z CA ~G(N,μ), so the false alarm probability can be derived and the detection probability The expression is:

[0067]

[0068]

[0069] In equations (9) and (10), n is the number of reference windows. It can be seen that the detection probability is independent of μ, that is, the CA-CFAR detector is a constant false alarm. The detection process of CA-CFAR can be summarized as follows: Given the false alarm probability of the system and the reference window length, the multiplication factor T can be obtained. CA The value of , and then substituting it into (10), we get the detection probability of the system.

[0070] The GO-CFAR constant false alarm detection is a mean-type CFAR algorithm developed to deal with the increase of false alarms in the clutter edge environment of CA-CFAR. Its background clutter power reference value Z GO The calculation formula is as follows

[0071] Z GO =max(A,B) (11)

[0072] In formula (11), A is the mean of the first half sliding window of the unit to be detected, and B is the mean of the second half sliding window of the unit to be detected. And the threshold factor T GO The relationship is as follows

[0073]

[0074] The mean-selected small (SO-CFAR) constant false alarm detection is a mean-type CFAR algorithm developed to deal with the increase in missed reports of CA-CFAR in a multi-target environment. The background clutter power reference value Z SO The calculation formula is as follows

[0075] Z SO =min(A,B) (13)

[0076] False Alarm Probability of SO-CFAR And the threshold factor T SO The relationship is as follows

[0077]

[0078] The process of the ACMLD algorithm is as follows:

[0079] 1) First, the sample q in the reference unit (1) ,q (2) …q (N) Sort from smallest to largest:

[0080] q (1) ≤q (2) ≤…≤q (N) (15)

[0081] Set an initial sample size k = 10;

[0082] 2) Compare q (k+1) and threshold T k s k Size:

[0083] If q (k+1) <T k s k , then q (k+1) ,…,q (N) does not contain interference targets, let k=k+1, go to 2).

[0084] If q (k+1) >T k s k , then q (k+1) ,…,q (N) Contains the interference target, go to 3).

[0085] in Each step corresponds to T k It can be calculated by the following formula:

[0086]

[0087] 3) The sample q (k+1) ,…,q (N) Delete, take q (1) ,…,q (k) For estimating background noise power:

[0088]

[0089] It can be seen that when the background environment is uniform clutter, CA-CFAR has the best detection performance. In a multi-target environment, the ACMLD-CFAR detection method has better performance. Selecting SO-CFAR in the low-power area of ​​the reverberation edge can effectively avoid the phenomenon of missed detection of targets at the edge. Using the GO-CFAR detection method in the high-power area of ​​the reverberation edge can reduce the false alarm of the system.

[0090] This embodiment provides the specific background clutter type to be used in the best detection method, so as to select a suitable CFAR detector.

[0091] Embodiment 5: This embodiment further limits the novel adaptive sonar target detection method based on neural network described in embodiment 1. In this embodiment, the step of classifying the sonar target signal according to the reference unit and using the neural network is further limited, specifically including: inputting the reference unit into the clutter background recognition classification network.

[0092] It should be noted that, in a more complex background environment, it is difficult to obtain more robust detection performance using only one constant false alarm detection method. If the type of environmental background in which the target is located can be first identified through a neural network algorithm, and then the appropriate CFAR detection method is adaptively adopted for different backgrounds, the performance can definitely be significantly improved. Therefore, this embodiment proposes a new sonar signal detector based on a neural network algorithm, with CA-CFAR, GO-CFAR, SO-CFAR and ACMLD-CFAR as optional options. Then, a clutter background recognition and classification network is selected as the neural network in this embodiment.

[0093] BP neural network is a multi-layer feedforward neural network trained according to the error back propagation algorithm. Assume that the input of the i-th unit in the k-th layer is:

[0094]

[0095] in is the connection weight between the i-th unit in layer k and the j-th unit in layer k-1; N k-1 is the number of neurons in the k-1th layer; is the threshold bias of the i-th unit in the k-th layer. Thus, its output is

[0096]

[0097] Where g(·) is the activation function. When k = 0,

[0098] Assume that the data sample set is {(x1,o1),(x2,o2),…,(x N ,o N )}, where x i is the input of training data, o N is the expected output of the training data. The error function of the network is:

[0099]

[0100] like Figure 4 As shown in the figure, the detector is mainly composed of two parts. The first part is to identify the background environmental clutter, and the second part is to select the CFAR detection method corresponding to the background. N These are N reference units obtained after processing the sonar echo signal. K groups of such reference units are sent to the recognition network, and finally the background recognition result is output. Each result is represented by a feature quantity. For example, the corresponding relationship between the feature quantity and the CFAR detector is shown in Table 1:

[0101] Table 1 Selection of detection methods corresponding to feature quantities

[0102] Feature quantity Background clutter type CFAR detection method 1000 Uniform background environment CA-CFAR 0100 Clutter edge high power area GO-CFAR 0010 Clutter edge low power area SO-CFAR 0001 Multi-target environment ACMLD-CFAR

[0103] Implementation method six: This implementation method further limits the novel adaptive sonar target detection method based on neural network described in implementation method five. In this implementation method, the clutter background recognition and classification network is further limited, specifically including: the clutter background recognition and classification network adopts the Levenberg-Marquardt algorithm for feedback training.

[0104] In this embodiment, the feedback mechanism of the BP neural network is as follows: Figure 3 As shown, for the LM feedback algorithm, the adjustment function of the weight and threshold bias is:

[0105] Δw k =[J T (w k )J(w k )+σI] -1 J T (w k )e(w k ) (twenty one)

[0106] Δb k =[J T (b k )J(b k )+σI] -1 J T (b k )e(b k ) (twenty two)

[0107] Where I is the identity matrix; σ is the user-defined learning rate; e(w) and e(b) represent the errors of weight w and threshold bias b respectively; J T (w k )e(w k ) and J T (b k )e(b k ) is the gradient; J(x) is the Jacobian matrix:

[0108]

[0109] This embodiment provides a method for feedback training in a neural network, specifically, feedback training is performed on the error between the actual output and the ideal output of the neural network.

[0110] Embodiment 7: This embodiment further limits the novel adaptive sonar target detection method based on a neural network described in Embodiment 1. In this embodiment, the step of selecting the best detection method according to the background clutter type of the detection unit is further limited, specifically including:

[0111] If the background clutter type is a uniform background environment, select the CA-CFAR detector;

[0112] If the background clutter type is a high power area at the edge of the clutter environment, select the GO-CFAR detector;

[0113] If the background clutter type is a low-power area at the edge of the clutter, a SO-CFAR detector is selected;

[0114] If the background clutter type is a multi-target background environment, the ACMLD-CFAR detector is selected.

[0115] For different sonar target echoes, CFAR detectors with better performance are selected. Among them, the CA-CFAR detector performs best in uniform background noise. The GO-CFAR detector is used in high-power areas with clutter edges to avoid false alarms. The GO-CFAR detector is used in low-power areas with clutter edges to avoid missed alarms. For multi-target environments, ACMLD-CFAR has better performance.

[0116] The technical solution of the present invention mainly includes:

[0117] Firstly, the received sonar target echo is input into the clutter background recognition and classification network, and the Levenberg-Marquardt (LM) algorithm is used for feedback training to classify the signal into four types: uniform background environment, high-power area at the edge of clutter environment, low-power area at the edge of clutter, and multi-target background environment.

[0118] For different sonar target echoes, CFAR detectors with better performance are selected. Among them, the CA-CFAR detector performs best in uniform background noise. The GO-CFAR detector is used in high-power areas with clutter edges to avoid false alarms. The GO-CFAR detector is used in low-power areas with clutter edges to avoid missed alarms. For multi-target environments, ACMLD-CFAR has better performance.

[0119] Sonar target detection has the characteristics of large data volume and complex reverberation background. In a complex background environment, it is difficult to obtain a more robust detection performance using only one constant false alarm detection method. In order to efficiently and accurately realize the detection of sonar targets, a new ANN-CFAR signal detector based on the neural network algorithm is proposed, which uses the characteristics of the BP neural network optimized by the Levenberg-Marquardt (LM) algorithm to classify and identify the input signal, and CA-CFAR, GO-CFAR, SO-CFAR and ACMLD-CFAR are optional. The method first inputs the received sonar target echo into the clutter background recognition classification network, and divides the signal into four types through feedback training: uniform background environment, high power area of ​​clutter edge environment, low power area of ​​clutter edge, and multi-target background environment. For different sonar target echoes, the CFAR detector with the best detection performance is selected in a targeted manner. This method has strong adaptability to complex background environments, and the detection performance has been significantly improved.

Claims

1. A novel adaptive sonar target detection method based on neural network, characterized in that: The method comprises: Perform matched filtering on the received sonar target signal; Performing detection processing on the filtered sonar target signal to obtain a detection unit and a reference unit; According to the reference unit, the sonar target signal is classified by using a neural network to obtain the background clutter type of the detection unit; Selecting the best detection method according to the background clutter type of the detection unit; Using the optimal detection method, the sonar target is detected; The classifying the sonar target signal by using a neural network specifically includes: inputting the reference unit into a clutter background recognition classification network, wherein the clutter background recognition classification network is feedback trained by using a Levenberg-Marquardt algorithm; The adjustment function of weight and threshold bias is: Δw k =[J T (w k )J(w k )+σI] -1 J T (w k )e(w k ) Δb k =[J T (b k )J(b k )+σI] -1 J T (b k )e(b k ) Where I is the unit matrix; σ is the user-defined learning rate; e(w) and e(b) represent the errors of weight w and threshold bias b respectively; J T (w k )e(w k ) and J T (b k )e(b k ) is the gradient; J(x) is the Jacobian matrix: The sonar target signal is divided into four types: uniform background environment, high power area at the edge of clutter environment, low power area at the edge of clutter environment and multi-target background environment; If the background clutter type is a uniform background environment, select the CA-CFAR detector; If the background clutter type is a high power area at the edge of the clutter environment, a GO-CFAR detector is selected; If the background clutter type is a low-power area at the edge of the clutter, a SO-CFAR detector is selected; If the background clutter type is a multi-target background environment, the ACMLD-CFAR detector is selected.

2. The novel adaptive sonar target detection method based on neural network according to claim 1 is characterized in that: The detection processing method includes: square law detection or linear detection.

3. The novel adaptive sonar target detection method based on neural network according to claim 1 is characterized in that: The acquisition method of the detection unit specifically includes: According to the filtered sonar target signal, obtaining an in-phase component and an orthogonal component of the sonar target signal; The detection unit is acquired according to the in-phase component and the orthogonal component.

4. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor runs the computer program stored in the memory, the steps of the method according to any one of claims 1 to 3 are performed.

Images