A Non-uniform Quantization Neural Network Fusion Detection Method for Underwater Weak Signals

Abstract

The present invention provides a non-uniform quantization neural network fusion detection method for underwater weak signals. Sonar is used to collect acoustic signals in the ocean. The signals received by the sonar are preprocessed by filtering and noise reduction. The measure value of the signal sample is calculated, and a non-uniform quantization neural network measure fusion model is constructed. The measure is input into the trained neural network, and the network output is calculated through the parameter weight matrix as the fusion measure representation and a decision is made. The present invention effectively improves the detection performance of weak signals under low signal-to-noise ratio conditions, and can approximate the optimal matched filtering detection performance under ideal conditions at a certain false alarm probability. Under the Neyman-Pearson criterion, by designing a measure fusion method of a non-uniform quantization neural network, a fusion measure is constructed as a test statistic to achieve effective detection of weak signals in water under strong background noise.

Description

Technical Field

[0001] The present invention relates to the field of signal detection, and in particular to a method for detecting weak underwater signals. Background Art

[0002] In order to respond to the new opportunities and challenges brought by the "Ocean Century" and further revitalize the ocean cause, in accordance with the strategic thinking of "building a strong ocean country", efforts should be made to develop marine science and technology and improve China's coastal defense strength. Therefore, researching advanced passive detection methods for weak signals has great research value and practical significance for detecting and identifying targets in water at a long distance.

[0003] The detection of ships at a long distance is mainly achieved by detecting the radiated noise of ships. The radiated noise of ships is a spectrum composed of special line spectra and continuous spectra. The detection of ship radiated noise can be realized by extracting the characteristics of the line spectra and continuous spectra. Traditional weak signal detection methods such as energy detection still have insufficient detection performance for ship radiated noise under low signal-to-noise ratios, and it is difficult to meet the growing requirements for detection distance and accuracy.

[0004] However, a neural network can use a complex network structure and a non-linear activation function to deeply characterize the features of a large amount of data and fit a more complex mapping relationship between the data. By non-linearly fusing various characteristic measures of ships, effective detection of weak underwater signals can be achieved. Summary of the Invention

[0005] In order to overcome the deficiencies of the prior art, the present invention provides a non-uniform quantization neural network fusion detection method for weak underwater signals. Under the Neyman-Pearson criterion, by designing a measure fusion method for a non-uniform quantization neural network, a fusion measure is constructed as a test statistic to achieve effective detection of weak underwater signals in strong background noise.

[0006] The specific steps of the technical solution adopted by the present invention to solve its technical problems are as follows:

[0007] First step: Use a sonar to collect acoustic signals in the ocean, denoted as g(t), which is the input signal; the input signal g(t) contains a mixture of target feature signals and noise signals, that is

[0008] g(t) = s(t) + n(t) (1)

[0009] where the target feature signal s(t) = A cos(2πf0t), A is the amplitude of the input signal, f0 is the frequency of the input signal, and n(t) is the ocean background noise signal;

[0010] Second step: Pass the signal g(t) received by the sonar in the first step through a band-pass filter for preprocessing of filtering and noise reduction;

[0011] Step 3: Taking the received signal-to-noise ratio SNR as the indication degree, non-uniformly divide the SNR interval into N intervals. The lower the SNR, the finer the division of the interval. For example, the interval from -30 dB to 0 dB can be non-uniformly divided into 15 intervals, obtain the noise-containing signal samples in different intervals, define the pure noise signal sample as belonging to the (N + 1)-th interval, and obtain (N + 1) kinds of signal samples;

[0012] Step 4: Calculate the R kinds of measurement values of the (N + 1) kinds of signal samples. The measurement values include spectral peak signal-to-noise ratio, spectral kurtosis, spectral correlation coefficient, mean square error, and smoothness, which are used as the training samples of the neural network, and use the interval number to which the signal sample belongs as the expected label;

[0013] Step 5: Construct a non-uniform quantization neural network measurement fusion model. Taking the R kinds of measurement values as the network input, construct a K-layer network, define the target output y i corresponding to the (N + 1) kinds of SNR intervals of the signal samples, i = 1, 2,..., N + 1. The size of the weight matrix of each layer of the network is [R*S1], [S1*S2],..., [S j-1 *S j ,..., [S K-1 *S K , [S K *(N + 1)], and the size of the bias matrix is [1*S1], [1*S2],..., [1*S j ,..., [1*S K , [1*(N + 1)], where Sj is the number of neurons corresponding to each layer of the network, j = 1, 2,..., K; design the network activation function as tanh, the network loss function as the cross-entropy loss function, and iteratively train to obtain the trained neural network, and save the parameter weight matrix W corresponding to the trained neural network;

[0014] Step 6: Input the R kinds of measurements of the signal to be detected into the trained neural network, and calculate the y to which the network output belongs through the parameter weight matrix W i as the fusion measurement representation, and use y i as the test statistic, and make a decision on the presence or absence of the target according to the threshold value under a specific false alarm probability. If the fusion measurement value is greater than or equal to the threshold value, it is judged as no target. If the fusion measurement value is less than the threshold value, it is judged as having a target.

[0015] In the second step, the bandwidth is 60 Hz, the maximum attenuation in the passband is 3 dB, and the minimum attenuation in the stopband is 40 dB.

[0016] The calculation formula of the spectral peak signal-to-noise ratio is:

[0017]

[0018] N is the number of sampling points for performing discrete Fourier transform on the received signal g(t), X(j) is the amplitude spectrum sequence of g(t), X(j0) represents the spectral line with the largest amplitude, i.e., the spectral peak, j is the subscript parameter, and j0 corresponds to the frequency of the spectral line.

[0019] The calculation formula for the spectral correlation coefficient is:

[0020]

[0021] and are the average values of X(j) and Y(j) respectively, and Y(j) is the amplitude spectrum sequence of s(t).

[0022] The calculation formula for the spectral kurtosis is:

[0023]

[0024] PX(j) is the energy spectrum sequence of g(t), is the average value of the energy spectrum sequence.

[0025] The calculation formula for the mean square error is:

[0026]

[0027] x[n] is the time-discrete sequence of g(t), y[n] is the time-discrete sequence of s(t), and M is the number of points in the time-discrete sequence.

[0028] The calculation formula for the smoothness is:

[0029]

[0030] where x[n + 1] and y[n + 1] are the next data points of x[n] and y[n] respectively.

[0031] The calculation formula for the tanh activation function is:

[0032]

[0033] x is the value of each neuron, and e is the base of the natural logarithm.

[0034] The calculation formula for the cross-entropy loss function is

[0035]

[0036] q(j) is the normalized value in the network output matrix of each sample indexed by the sample label value.

[0037] The beneficial effects of the present invention are as follows. Since a neural network is used to fuse multiple characterization measures of signals, and the fusion measure designed by the non-uniform quantization neural network is used as the test statistic, the detection performance of weak signals is much better than that of traditional detectors that use a single measure (such as energy) for detection. It effectively improves the detection performance of weak signals under low signal-to-noise ratio conditions and can approach the optimal matched filtering detection performance under ideal conditions at a certain false alarm probability. The selection of the fusion measure includes but is not limited to the following five measures: spectral peak signal-to-noise ratio, spectral kurtosis, spectral correlation coefficient, mean square error, and smoothness. Description of the Drawings

[0038] Figure 1 It is a non-uniform quantization neural network model diagram constructed by the present invention.

[0039] Figure 2 It is a principle block diagram of the present invention.

[0040] Figure 3 It is a curve of the detection performance of the present invention varying with the signal-to-noise ratio.

[0041] Figure 4 It is a receiver operating characteristic (ROC) curve of the detection effect of the present invention. Detailed Embodiment

[0042] The present invention will be further described below in conjunction with the drawings and embodiments.

[0043] The first step: Use a sonar to collect acoustic signals in the ocean, denoted as g(t), which is the input signal. The input signal contains a mixture of single-frequency line spectrum signals and noise signals, that is

[0044] g(t) = s(t) + n(t) (1)

[0045] where s(t) = A cos(2πf0t), A is the amplitude of the input signal, f0 is the frequency of the input signal, and n(t) is the ocean background noise signal;

[0046] The second step: Pass the signal g(t) received by the sonar in the first step through a Butterworth band-pass filter for filtering and noise reduction. It is recommended that the bandwidth be 60 Hz, the maximum attenuation in the passband be 3 dB, and the minimum attenuation in the stopband be 40 dB;

[0047] The third step: Using the received signal-to-noise ratio SNR as an indicator, non-uniformly divide the signal-to-noise ratio interval into N intervals. The lower the signal-to-noise ratio, the finer the division. For example, the interval from -30 dB to 0 dB can be non-uniformly divided into 15 intervals. Obtain noise signal samples in different intervals, and define the pure noise signal sample as the (N + 1)-th interval to obtain (N + 1) kinds of signal samples;

[0048] An example of the method for obtaining training samples is as follows: At high signal-to-noise ratio, from -1 dB to -21 dB, 1000 signal samples are obtained for every 5 dB reduction, totaling 5000 samples. At low signal-to-noise ratio, from -25.5 dB to -30 dB, 1000 signal samples are obtained for every 0.5 dB reduction, totaling 10000 samples. 1000 signal samples are still obtained under pure noise conditions.

[0049] Step 4: For example, calculate various measures of the above-mentioned 16000 signal samples in total to form the input matrix of the neural network. Taking five measures, namely spectral peak signal-to-noise ratio, spectral correlation coefficient, spectral kurtosis, smoothness, and mean square error, as an example, a network input matrix of size [16000*5] is formed, where 16000 represents the number of signal samples and 5 represents the above five measures.

[0050] Among them, the calculation formula for the spectral peak signal-to-noise ratio is

[0051]

[0052] N is the number of sampling points for the discrete Fourier transform of the received signal g(t), X(j) is the amplitude spectrum sequence of g(t), X(j0) represents the spectral line with the largest amplitude, that is, the spectral peak, j is the subscript parameter, and j0 corresponds to the frequency of the spectral line;

[0053] The calculation formula for the spectral correlation coefficient is:

[0054]

[0055] and are the average values of X(j) and Y(j) respectively, and Y(j) is the amplitude spectrum sequence of s(t);

[0056] The calculation formula for the spectral kurtosis is:

[0057]

[0058] PX(j) is the energy spectrum sequence of g(t), is the average value of the energy spectrum sequence;

[0059] The calculation formula for the mean square error is:

[0060]

[0061] x[n] is the time-discrete sequence of g(t), y[n] is the time-discrete sequence of s(t), and M is the number of points in the time-discrete sequence;

[0062] The calculation formula for the smoothness is:

[0063]

[0064] Among them, x[n + 1] and y[n + 1] are the next data points of x[n] and y[n] respectively;

[0065] Step 5: Construct a non-uniform quantization neural network measure fusion model. Using R measure values as the network input, construct a K-layer network, and define the target output y i For the N + 1 signal-to-noise ratio intervals corresponding to the signal samples, i = 1, 2,..., N + 1, the sizes of the weight matrices of each layer of the network are [R * S1], [S1 * S2],..., [S K-1 * S K , [S K *(N + 1)], and the sizes of the bias matrices are [1 * S1], [1 * S2],..., [1 * S K , [1 * (N + 1)], where S j is the number of neurons corresponding to each layer of the network, j = 1, 2,..., K; Design the network activation function (such as tanh) and the network loss function (such as cross-entropy loss function), and iteratively train to obtain the parameter weight matrix W of each specific layer of the neural network;

[0066] Construct a non-uniform quantization neural network model as Figure 1 shown. The following suggestions are given for the number of network layers and the size of each layer of the network: Construct an example of a 3-layer network. The sizes of the weight matrices of each layer of the network are [5 * 8], [8 * 12], [12 * 16], and the sizes of the bias matrices are [1 * 8], [1 * 12], [1 * 16] in sequence. The activation function of each layer of the network is tanh, and the loss function of the network is set as the cross-entropy loss function.

[0067] Among them, the calculation formula of the tanh activation function is:

[0068]

[0069] x is the value of each neuron, and e is the base of the natural logarithm, approximately equal to 2.71828.

[0070] The calculation formula of the cross-entropy loss function is

[0071]

[0072] q(j) is the normalized value indexed by the sample label value in the network output matrix of each sample.

[0073] With the minimum gradient descent criterion, continuously backpropagate to train the neural network; After backpropagating several times, when the error is less than 10 -4 the training of the neural network is completed, and the parameter weight matrix W of the network is saved. The following suggestions are given for the specific values of each layer of the weight matrix:

[0074] The first layer:

[0075] w1

[0076] = [[0.290, -0.505, 2.668, -1.334, -3.060, -1.904, 2.779, -1.513] [1.343, -0.448, -1.281, -1.396, 0.895, 3.651, -0.306, -0.608] [1.428, -1.356, -0.190, 1.461, 0.021, 1.066, 0.781, -1.099] [0.575, 0.838, -3.010, 0.624, 1.967, -0.270, -1.472, -0.406] [0.121, 1.949, -0.284, 1.060, 1.341, -4.471, -1.233, 1.329]]

[0077] The second layer:

[0078] w2

[0079] =[[-1.038, -1.412, 0.647, 0.256, -0.629, 0.554, -0.459, -1.195, -0.431, 0.397, 1.338, 0.896] [0.214, 2.215, 1.625, 1.008, 0.498, -2.105, 1.048, -0.129, -0.334, 1.568, -1.010, -0.496] [2.246, 0.902, 0.075, 0.925, 1.428, 0.317, 0.252, 1.039, 1.772, -1.441, -0.623, -0.226] [0.639, -0.756, 0.935, -0.075, 0.734, 0.216, 0.489, 0.421, 0.228, 0.727, 0.453, 1.008] [-0.044, 0.143, 1.655, -1.358, -0.364, 0.032, 0.998, -0.757, -1.075, 0.352, 0.438, 0.870] [1.070, 1.253, -0.144, -2.144, 0.610, 0.271, -2.420, 0.037, 0.377, -1.438, 1.110, -0.039] [-1.186, -0.883, -1.245, -0.513, -0.306, 1.053, 0.604, 0.021, 0.602, -1.914, -0.427, -0.340] [-0.314, -0.237, -0.200, 0.299, 1.560, 0.534, 0.621, -0.260, -0.667, 1.297, 1.254, 0.269]

[0080] The third layer:

[0081] w3

[0082] ＝[[-0.508，0.498，0.720，0.222，-0.372，1.594，0.440，-1.917，-1.109，-0.523，-1.066，-1.301，-1.208，0.429，-0.445，-0.409] [-1.907，-0.519，0.781，-0.734，0.058，0.448，-1.889，-0.224，-0.544，1.411，1.213， 1.253，-1.307，-0.252，1.175，0.403] [-2.531，-1.132，0.608，-0.018，0.710，1.932，-0.920，-0.084，1.271，-0.407，-0.333， 0.077，0.517，-0.408，1.617，1.532，] [0.335，-0.498，-0.789，-0.897，1.130，0.021，0.242，-0.455，-0.121，1.828，-0.892，0.052， 0.091，-0.318，-0.851，0.556] [-0.701，-1.437，0.994，-1.719，0.519，-0.852，0.266，1.797，2.097，-0.055，-0.418，1.034，-1.054，-0.812，-0.436，-1.114] [0.387，0.300，0.030，0.381，1.831，0.022，-0.095，-1.346，0.235，-0.230，-0.889，-1.283， -0.209，0.311，-0.240，-1.300] [1.764，-1.049，0.073，-1.017，-0.683，-0.486，-0.621，-0.259，0.697，-0.074，-1.279， -0.265，-1.256，-0.149，0.213，0.812] [0.704，1.452，-0.094，0.917，-0.089，-0.850，-0.594，0.012，-0.995，1.889，-1.213， 0.412，-1.757，0.094，-0.458，1.420] [-1.621，1.385，-1.027，-0.363，-0.518，-0.758，-1.228，-0.620，-1.515，0.572，-1.769， 0.148，-0.152，-1.960，-1.225，0.712] [-0.606，-1.750，-1.809，-0.[467, 0.773, -0.243, 1.012, -1.512, -1.311, -0.702, -0.155, -0.973, -1.233, -1.000, 2.723, 0.011] [0.225, 0.213, -0.964, 0.547, -0.753, 1.399, 0.993, 0.118, 0.371, 0.650, 0.433, -0.732, -0.009, 2.260, 0.929, 0.713] [-0.241, 0.572, 0.230, 1.119, -0.180, 0.077, 0.265, 0.891, 2.296, 0.913, -0.238, -0.025, 0.328, -0.904, 0.147, -0.104]].

[0083] The following suggestions are given for the specific values of each layer of the bias matrix:

[0084] The first layer:

[0085] b1 = [0.279, -1.567, 0.323, -0.195, -0.112, 0.218, 0.712, -0.668]

[0086] The second layer:

[0087] b2 = [0.135, -0.123, 0.083, -0.075, -0.199, 0.205, -0.199, -0.396, -0.280, -0.145, 0.078, 0.311]

[0088] The third layer:

[0089] b3 = [0.065, 0.005, -0.321, 0.618, -0.319, 0.212, 0.179, 0.152, 0.257, -0.245, 0.059, -0.109, 0.026, 0.186, -0.002, -0.001]

[0090] Step 6: Input the M measures of the signal to be detected into the trained neural network, and directly calculate the y to which the network output belongs through the parameter weight matrix W i As the fusion measure representation, with y i As the test statistic, make a decision on the presence or absence of the target according to the threshold value under a specific false alarm probability. If the fusion measure value is greater than or equal to the threshold value, it is judged as no target; if the fusion measure value is less than the threshold value, it is judged as having a target.

[0091] Figure 3 is the curve of the detection performance of the present invention varying with the signal-to-noise ratio,Figure 4 is the receiver operating characteristic (ROC) curve of the detection effect of the present invention, where P F is the false alarm probability, and P D is the detection probability.

Claims

1. A non-uniform quantization neural network fusion detection method for underwater weak signals, characterized in that It includes the following steps: The first step: Use sonar to collect acoustic signals in the ocean, denoted as g(t), which is the input signal; the input signal g(t) contains a mixture of target feature signals and noise signals, that is g(t) = s(t) + n(t) (1) where the target feature signal s(t) = Acos(2πf0t), A is the amplitude of the input signal, f0 is the frequency of the input signal, and n(t) is the ocean background noise signal; The second step: Pass the signal g(t) received by the sonar in the first step through a band-pass filter for preprocessing of filtering and noise reduction; The third step: Using the received signal-to-noise ratio SNR as an indicator, non-uniformly divide the signal-to-noise ratio interval into N intervals. The lower the signal-to-noise ratio, the more detailed the division of the interval. For example, the interval from -30dB to 0dB can be non-uniformly divided into 15 intervals, obtain noise signal samples in different intervals, define the pure noise signal sample as the (N + 1)-th interval, and get (N + 1) kinds of signal samples; The fourth step: Calculate R measurement values of the (N + 1) kinds of signal samples. The measurement values include spectral peak signal-to-noise ratio, spectral kurtosis, spectral correlation coefficient, mean square error, and smoothness, as the training samples of the neural network, and use the interval number to which the signal sample belongs as the expected label; Step 5: Construct a non-uniform quantization neural network measurement fusion model. Using R measurement values as the network input, construct a K-layer network, and define the target output y i For the N+1 signal-to-noise ratio intervals corresponding to the signal samples, i = 1, 2,..., N+1, the sizes of the weight matrices of each layer of the network are [R*S1], [S1*S2],..., [S j-1 *S j ,..., [S K-1 *S K , [S K *(N+1)], and the sizes of the bias matrices are [1*S1], [1*S2],..., [1*S j ,..., [1*S K , [1*(N+1)], where S j is the number of neurons corresponding to each layer of the network, j = 1, 2,..., K; design the network activation function as tanh and the network loss function as the cross-entropy loss function, and iteratively train to obtain a trained neural network, and save the parameter weight matrix W corresponding to the trained neural network Step 6: Input the R measures of the signal to be detected into the trained neural network, and calculate the y to which the network output belongs through the parameter weight matrix W i As the fusion measure representation, use y i As the test statistic, make a decision on the presence or absence of a target according to the threshold value under a specific false alarm probability. If the fusion measure value is greater than or equal to the threshold value, it is judged that there is no target. If the fusion measure value is less than the threshold value, it is judged that there is a target.

2. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, characterized in that: In the second step, the bandwidth is 60Hz, the maximum attenuation in the passband is 3dB, and the minimum attenuation in the stopband is 40dB.

3. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, characterized in that: The calculation formula for the spectral peak signal-to-noise ratio is: N is the number of sampling points for the discrete Fourier transform of the received signal g(t), X(j) is the amplitude spectrum sequence of g(t), X(j0) represents the spectral line with the largest amplitude, that is, the spectral peak, j is the subscript parameter, and j0 corresponds to the frequency of the spectral line.

4. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, characterized in that: The calculation formula for the spectral correlation coefficient is: and are the average values of X(j) and Y(j) respectively, where Y(j) is the amplitude spectrum sequence of s(t).

5. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, characterized in that: The calculation formula for the spectral kurtosis is: PX(j) is the energy spectrum sequence of g(t), which is the average value of the energy spectrum sequence.

6. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, characterized in that: The calculation formula for the mean square error is: x[n] is the time-discrete sequence of g(t), y[n] is the time-discrete sequence of s(t), and M is the number of points in the time-discrete sequence.

7. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, characterized in that: The calculation formula for the smoothness is: where x[n + 1] and y[n + 1] are the next data points of x[n] and y[n] respectively.

8. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, characterized in that: The calculation formula for the tanh activation function is: x is the value of each neuron, and e is the base of the natural logarithm.

9. The non-uniform quantization neural network fusion detection method for underwater weak signals according to claim 1, wherein: The calculation formula of the cross-entropy loss function is q(j) is the normalized value indexed by the sample label value in the network output matrix of each sample.

Images