Single-frequency feature extraction method based on one-dimensional convolutional neural network and enhanced discrete feature loss function

Abstract

The invention provides a single-frequency feature extraction method based on a one-dimensional convolutional neural network and an enhanced discrete feature loss function. A feature extraction model sensitive to single-frequency features is obtained through training by designing a one-dimensional convolutional neural network structure model and utilizing simulation data and training by designing a discrete feature loss function. The method is used for power spectrum and LOFAR spectrogram single-frequency feature extraction, and a good effect is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of underwater acoustic target parameter estimation and artificial intelligence technology, and mainly relates to a single-frequency feature extraction method based on a one-dimensional convolutional neural network and an enhanced discrete feature loss function. Background Art

[0002] The single-frequency feature of the underwater acoustic target radiation noise power spectrum / LOFAR spectrum is an important feature in underwater acoustic target recognition. The single-frequency feature is often related to the operation of equipment such as the main engine and auxiliary engine of the underwater acoustic target, and belongs to the inherent feature information of the underwater acoustic target, which can provide important information for target recognition.

[0003] However, for the current method of extracting the target single-frequency feature by a simple threshold method, the target radiation noise is affected by the transmission of the ocean acoustic channel and the interference of other targets, and the extracted single-frequency feature has a high false alarm rate and poor stability. At the same time, for the single-frequency feature extraction method using a complex neural network, the computational cost is large and it is difficult to meet the requirements of real-time processing. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies existing in the prior art, and provide a single-frequency feature extraction method based on a one-dimensional convolutional neural network and an enhanced discrete feature loss function.

[0005] The purpose of the present invention is achieved by the following technical solutions. A single-frequency feature extraction method based on a one-dimensional convolutional neural network and an enhanced discrete feature loss function includes the following steps:

[0006] S1. Construct a network model for single-frequency feature extraction based on a one-dimensional convolutional neural network;

[0007] (1.1) Construct a power spectrum / LOFAR spectrum map input layer

[0008] Construct an input layer adapted to the input power spectrum / LOFAR spectrum map. The number of input layer nodes is greater than or equal to the number of frequency points of the spectrum map; when the number of input layer nodes is equal to the number of frequency points of the spectrum map, the node values are equal to the spectrum map frequency point values in sequence and are input into the network; when the number of input layer nodes is greater than the number of frequency points of the spectrum map, through interpolation, the spectrum map frequency points are expanded to be equal to the number of input layer nodes and are input into the network;

[0009] (1.2) Construct a multi-layer single-frequency feature extraction structure

[0010] Based on a 5-layer convolutional layer model, set the number of input sample channels to 1, reconstruct the convolutional operator, modify the longitudinal dimension to 1, and lengthen the transverse dimension. Each convolutional layer uses ReLU as the activation function;

[0011] The convolution kernel of the first convolutional layer is (1, 5), the stride is (1, 4), and the output result size is 1×(N输入 / 4);

[0012] The convolution kernel of the second convolutional layer is (1, 5), the stride is (1, 4), and the output result size is 1×(N 输入 / 16);

[0013] The convolution kernel of the third convolutional layer is (1, 5), the stride is (1, 4), and the output result size is 1×(N 输入 / 64);

[0014] The convolution kernel of the fourth convolutional layer is (1, 3), the stride is (1, 2), and the output result size is 1×(N 输入 / 128);

[0015] The convolution kernel of the fifth convolutional layer is (1, 3), the stride is (1, 2), and the output result size is 1×(N 输入 / 256);

[0016] (1.3) Construct an output layer adapted to the single - frequency feature label

[0017] The number of nodes in the output layer is equal to the number of input nodes N 输入 to ensure that the number of nodes in the output result has a corresponding relationship with the number of frequency points of the input power spectrum / LOFAR spectrum, so as to determine the frequency characteristics of the single - frequency feature;

[0018] The output layer uses the Sigmod activation function, and the calculation method is shown in the following formula:

[0019]

[0020] to make the output result distributed between (0, 1) and adapted to the discrete feature label;

[0021] S2. Use the simulation data for training, and train a feature extraction model sensitive to single - frequency features by designing a loss function that enhances discrete features;

[0022] S3. Apply the network model to process the target radiation noise and extract the target single - frequency feature.

[0023] The steps for generating the target single - frequency feature label are as follows: First, separate the continuous - spectrum feature and the single - frequency feature in the target power spectrum / LOFAR spectrum to obtain the single - frequency feature with the relative energy magnitude relationship. Subsequently, determine the corresponding position of the single - frequency feature according to the frequency band and the number of label points, and determine the label value at the corresponding position according to the relative energy magnitude relationship.

[0024] The method for calculating the loss function for enhancing discrete features is as follows: First, calculate the deviation between the label and the prediction result, and then use the label to enhance the deviation corresponding to the discrete part. Specifically, multiply the deviation corresponding to the non-zero positions of the label by the discrete feature weight coefficient to increase its proportion in the loss function calculation. The specific calculation method is as follows:

[0025]

[0026] Where y is the label vector, f(x) is the predicted single-frequency feature, N is the vector length, and b ∈ [0, 1) is the discrete feature weight coefficient; the higher the value of b, the higher the weight of the discrete feature, and the more sensitive the loss value is to the discrete feature.

[0027] Apply the network model to process the target radiated noise and extract the target single-frequency feature. The basic process is as follows:

[0028] (3.1) Generate a power spectrum / LOFAR spectrum of the underwater acoustic target at a certain resolution, determine the number of frequency bands N, and perform global normalization;

[0029] (3.2) Input the standard power spectrum / LOFAR spectrum obtained in (3.1) into the one-dimensional convolutional neural network for single-frequency feature extraction to obtain the corresponding single-frequency feature;

[0030] (3.3) Process the newly acquired data according to (3.1) to (3.2) to obtain the continuous result of target single-frequency feature extraction.

[0031] The beneficial effects of the present invention are as follows: The present invention provides a method for extracting single-frequency features based on a one-dimensional convolutional neural network and an enhanced discrete feature loss function. By designing a fully connected neural network adapted to the data and training with the enhanced discrete feature loss function, a neural network for extracting single-frequency features with relatively small computational overhead is obtained, which can accurately extract the single-frequency features contained in the target radiated noise and provide a basis for target recognition. The present invention is used for extracting single-frequency features from power spectra and LOFAR spectra, and good results are obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the following described drawings are only some embodiments of the present invention. For those skilled in the art or ordinary technicians, other drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 The figure shows a schematic diagram of a typical power spectrum / LOFAR spectrum of underwater acoustic target radiated noise.

[0034] Figure 2The following is a schematic diagram of the target single - frequency feature label generation process adopted by the present invention.‘

[0035] Figure 3 The following is a network model diagram of underwater acoustic target single - frequency feature extraction.

[0036] Figure 4 The following is a schematic diagram of the calculation process of the enhanced discrete feature loss function of the present invention. Detailed implementation manners

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0038] As Figure 1 The following is a schematic diagram of the typical power spectrum / LOFAR spectrum of underwater acoustic target radiated noise, which is mainly composed of continuous spectrum features and single - frequency features. The single - frequency features have concentrated energy, good recognition characteristics, and reflect the working state and inherent characteristics of underwater acoustic targets.

[0039] As Figure 2 The following is the target single - frequency feature label generation process adopted by the present invention. First, the continuous spectrum features and single - frequency features in the target power spectrum / LOFAR spectrum are separated to obtain single - frequency features with relative energy magnitude relationships. Subsequently, according to the frequency band and the number of label points, the corresponding positions of the single - frequency features are determined, and according to the relative energy magnitude relationships, the label values at the corresponding positions are determined.

[0040] The present invention provides a single - frequency feature extraction method based on a one - dimensional convolutional neural network and an enhanced discrete feature loss function. The specific steps are as follows:

[0041] (1) Based on a one - dimensional convolutional neural network, a network model for single - frequency feature extraction is constructed. The basic process is as follows:

[0042] (1.1) Construct a power spectrum / LOFAR spectrum map input layer:

[0043] Construct an input layer adapted to the input power spectrum / LOFAR spectrum map. To ensure that as much information as possible of the input power spectrum / LOFAR spectrum is preserved, the number of input layer nodes is greater than or equal to the number of frequency points of the spectrum map. When the number of input layer nodes is equal to the number of frequency points of the spectrum map, the node values are equal to the spectrum map frequency point values in sequence and are input into the network. When the number of input layer nodes is greater than the number of frequency points of the spectrum map, through interpolation, the number of spectrum map frequency points is expanded to be equal to the number of input layer nodes and then input into the network.

[0044] (1.2) Construct a multi - layer single - frequency feature extraction structure

[0045] It is constructed based on a 5-layer convolutional layer model. The number of input sample channels is set to 1, and the convolutional operator is reconstructed. Given the high frequency resolution of the one-dimensional power spectrum / LOFAR spectrogram, the longitudinal dimension is modified to 1, and the transverse dimension is moderately lengthened. Each convolutional layer uses ReLU as the activation function.

[0046] The convolutional kernel of the first convolutional layer is (1, 5), the stride is (1, 4), and the output result size is 1×(N_input / 4);

[0047] The convolutional kernel of the second convolutional layer is (1, 5), the stride is (1, 4), and the output result size is 1×(N_input / 16);

[0048] The convolutional kernel of the third convolutional layer is (1, 5), the stride is (1, 4), and the output result size is 1×(N_input / 64);

[0049] The convolutional kernel of the fourth convolutional layer is (1, 3), the stride is (1, 2), and the output result size is 1×(N_input / 128);

[0050] The convolutional kernel of the fifth convolutional layer is (1, 3), the stride is (1, 2), and the output result size is 1×(N_input / 256);

[0051] (1.3) Construct an output layer adapted to the single-frequency feature label

[0052] The number of nodes in the output layer is equal to the number of input nodes N_input. This ensures that the number of nodes in the output result has a corresponding relationship with the number of frequency points of the input power spectrum

[0053] / LOFAR spectrogram, thereby determining the frequency characteristics of the single-frequency feature.

[0054] The output layer uses the Sigmod activation function, and the calculation method is shown in the following formula:

[0055]

[0056] This makes the output result distributed between (0, 1), which is adapted to the discrete feature label;

[0057] Such as Figure 3The figure shows a network model diagram for extracting the single-frequency features of an underwater acoustic target, which mainly consists of an input layer, multiple layers of feature extraction, and an output layer. Among them, the number of nodes in the input layer is greater than or equal to the number of frequency points of the spectrogram, and all input information is saved. The multi-layer single-frequency feature extraction structure is constructed based on a 5-layer convolutional layer model. The number of channels of the input sample is set to 1, and the vertical size of the convolutional kernel is modified to 1. The number of nodes in the output layer is equal to the number of input nodes N. This ensures that the number of nodes in the output result has a corresponding relationship with the number of frequency points of the input power spectrogram / LOFAR spectrogram, thereby determining the frequency characteristics of the single-frequency features.

[0058] (2) Design an enhanced discrete feature loss function for training the single-frequency feature extraction network model

[0059] Figure 4 The figure shows the calculation process of the enhanced discrete feature loss function designed for this intellectual achievement. First, calculate the deviation between the label and the prediction result, and then use the label to enhance the deviation corresponding to the discrete part. Specifically, multiply the deviation corresponding to the non-zero positions of the label by the discrete feature weight coefficient, thereby increasing the proportion in the calculation of the loss function and enhancing the sensitivity of the loss function to the discrete features.

[0060] The result of single-frequency feature extraction is a 1D vector, and its length is proportional to the length and resolution of the corresponding frequency band of the power spectrum. Since the number of single-frequency features in the target radiated noise is generally several or dozens, most of the elements in the label vector corresponding to the single-frequency features are 0 (no single-frequency feature), and a small number of elements are greater than 0 (there is a single-frequency feature). In order to increase the proportion of the elements with single-frequency features in the loss value, enhance the sensitivity of the loss value to the single-frequency features, and improve the training efficiency of the neural network model, an enhanced discrete feature loss function is proposed, and the calculation method is as follows:

[0061]

[0062] Among them, y is the label vector, f(x) is the predicted single-frequency feature, N is the vector length, and b ∈ [0, 1) is the discrete feature weight coefficient. When the value of b is higher, the weight of the discrete feature is higher, and the loss value is more sensitive to the discrete feature.

[0063] (3) Apply the network model to process the target radiated noise and extract the target single-frequency features. The basic process is as follows.

[0064] (3.1) Generate a power spectrum / LOFAR spectrum of the underwater acoustic target at a certain resolution, determine the number of frequency bands N, and perform global normalization.

[0065] (3.2) Input the standard power spectrum / LOFAR spectrum obtained in (3.1) into the one-dimensional convolutional neural network for single-frequency feature extraction to obtain the corresponding single-frequency features.

[0066] (3.3) Process the newly acquired data according to (3.1) to (3.2) to obtain the continuous results of target single-frequency feature extraction.

[0067] As mentioned above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. A single frequency feature extraction method based on a one-dimensional convolutional neural network and an enhanced discrete feature loss function, characterized in that: The following steps are involved: S1. Construct a network model for single frequency feature extraction based on one-dimensional convolutional neural network; (1.1) Constructing the power spectrum / LOFAR spectrogram input layer Construct an input layer that is compatible with the input power spectrum / LOFAR spectrogram, and the number of input layer nodes is greater than or equal to the number of frequency points of the spectrogram; when the number of input layer nodes is equal to the number of frequency points of the spectrogram, the node values ​​are sequentially equal to the frequency points of the spectrogram and input into the network; when the number of input layer nodes is greater than the number of frequency points of the spectrogram, the frequency points of the spectrogram are expanded to be equal to the number of input layer nodes through interpolation and input into the network; (1.2) Constructing a multi-layer single-frequency feature extraction structure Based on the 5-layer convolutional layer model, the number of input sample channels is set to 1, the convolution operator is reconstructed, the vertical size is modified to 1, the horizontal size is lengthened, and each convolution layer uses ReLU as the activation function; The convolution kernel of the first convolution layer is (1, 5), the step size is (1, 4), and the output size is 1×(N 输入 / 4); The convolution kernel of the second convolution layer is (1, 5), the step size is (1, 4), and the output size is 1×(N 输入 / 16); The convolution kernel of the third convolution layer is (1, 5), the step size is (1, 4), and the output size is 1×(N 输入 / 64); The convolution kernel of the fourth convolution layer is (1, 3), the step size is (1, 2), and the output size is 1×(N 输入 / 128); The convolution kernel of the fifth convolution layer is (1, 3), the step size is (1, 2), and the output size is 1×(N 输入 / 256); (1.3) Construct an output layer that is compatible with the single-frequency feature label The number of output layer nodes and the number of input nodes N 输入 Equal to ensure that the number of nodes in the output result corresponds to the number of frequency points in the input power spectrum / LOFAR spectrum, thereby determining the frequency characteristics of the single-frequency feature; The output layer uses the Sigmod activation function, and the calculation method is shown in the following formula: Make the output result distributed between (0,1) and adapt to the discrete feature label; S2. Use simulation data for training, and design a loss function that enhances discrete features to obtain a feature extraction model that is sensitive to single-frequency features; S3. Use the network model to process the target radiation noise and extract the target single-frequency characteristics.

2. The single frequency feature extraction method based on one-dimensional convolutional neural network and enhanced discrete feature loss function according to claim 1, characterized in that: The target single-frequency feature label generation step is: first, the continuous spectrum feature and the single-frequency feature in the target power spectrum / LOFAR spectrum are separated to obtain the single-frequency feature by the relative energy size relationship, and then the corresponding position of the single-frequency feature is determined according to the frequency band and the number of label points, and the corresponding position label value is determined according to the relative energy size relationship.

3. The single frequency feature extraction method based on one-dimensional convolutional neural network and enhanced discrete feature loss function according to claim 2, characterized in that: The loss function calculation method for enhancing discrete features is: first calculate the deviation between the label and the prediction result, and then use the label to enhance the deviation corresponding to the discrete part. Specifically, the deviation corresponding to the non-zero position of the label is multiplied by the discrete feature weight coefficient, thereby increasing the proportion in the loss function calculation; the specific calculation method is as follows: Where y is the label vector, f(x) is the predicted single-frequency feature, N is the vector length, and b∈[0,1) is the discrete feature weight coefficient; when the b value is higher, the weight of the discrete feature is higher, and the loss value is more sensitive to the discrete feature.

4. The single frequency feature extraction method based on one-dimensional convolutional neural network and enhanced discrete feature loss function according to claim 3, characterized in that: The network model is used to process the target radiation noise and extract the target single-frequency characteristics. The basic process is as follows: (3.1) Generate a power spectrum / LOFAR spectrum of the underwater acoustic target at a certain resolution, determine the number of frequency bands N, and perform global normalization; (3.2) Input the standard power spectrum / LOFAR spectrum obtained in (3.1) into the single-frequency feature extraction one-dimensional convolutional neural network to obtain the corresponding single-frequency feature; (3.3) Process the newly acquired data according to (3.1) to (3.2) to obtain continuous results of target single-frequency feature extraction.

Images