A small sample number underwater target recognition method based on deep learning

Abstract

The application provides a small sample number underwater target recognition method based on deep learning, comprising the following steps: S1, modeling underwater target radiation noise, giving an underwater target radiation noise DEMON spectrum processing method, and pre-processing the DEMON spectrum; S2, processing the radiation noise, simulating different working states and motion states of the target, and extracting sample combinations to generate sample pairs for training, verification and testing of a neural network model; S3, designing a one-dimensional convolutional neural network-based twin network model, training parameters of the convolutional neural network, evaluating network performance by using a verification set containing different Doppler frequency offsets, different signal-to-noise ratios and different spectral line numbers, obtaining a convolutional neural network model, and using the model to calculate the "similarity" of two underwater target radiation noise samples, and further judging whether the two samples are the same target or not. The application has the beneficial effect that the small sample number underwater target recognition problem can be solved.

Description

Technical Field

[0001] This invention belongs to the field of underwater target recognition technology, and in particular relates to a method for underwater target recognition with a small sample size based on deep learning. Background Technology

[0002] Passive sonar is one of the main methods for detecting and monitoring underwater targets. Passive classification and identification of underwater targets is crucial for port and coastal surveillance, underwater acoustic countermeasures, and maritime operations, and is of great significance in seabed resource development and national defense. Radiated noise is the primary source of information for underwater target classification and identification. Radiated noise includes various sources such as mechanical noise, hydrodynamic noise, and propeller noise. How to extract information reflecting target characteristics from radiated noise and classify and identify them has always been a recognized challenge both domestically and internationally, and remains an urgent problem to be solved in the field of underwater target classification and identification.

[0003] There are two main challenges in underwater target identification: First, underwater target samples are scarce, and acquiring samples of underwater target radiated noise is very costly, with limited sample quantity and variety, which is detrimental to training the identification system. Second, the working and motion states of a target lead to a diversity of samples for the same target (different Doppler frequency offsets, different signal-to-noise ratios, and different interference conditions), while the collected samples may only contain a limited number of states, making it difficult to establish a complete sample library for the same target, thus causing difficulties in target identification. Summary of the Invention

[0004] In view of this, the present invention aims to propose a deep learning-based method for identifying underwater targets with a small sample size, in order to solve the problem of identifying underwater targets with a small sample size when underwater target radiation noise samples are diverse.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0006] A deep learning-based method for underwater target identification with a small sample size includes the following steps:

[0007] S1. Model underwater target radiated noise, give a method for processing underwater target radiated noise DEMON spectrum, find the maximum value between channels of multi-channel DEMON spectrum to obtain single-channel DEMON spectrum, and perform mean subtraction operation on single-channel DEMON spectrum as a sample for training and validation of convolutional neural network;

[0008] S2. Process the underwater target radiated noise, simulate different working and motion states of the underwater target radiated noise, add Doppler frequency offset, signal-to-noise ratio, and interference spectrum lines, generate the training dataset, validation dataset, and test dataset required for the convolutional neural network, and combine them to generate sample pairs, which are used for training, validation, and testing of the convolutional neural network model.

[0009] S3. Design a convolutional neural network model:

[0010] The parameters of the convolutional neural network were trained using the original DEMON spectrum of underwater target radiated noise. The network performance was evaluated using a validation set with different Doppler frequency offsets, different signal-to-noise ratios, and different numbers of spectral lines to obtain the convolutional neural network model. The convolutional neural network model was used to calculate the similarity function of two underwater target radiated noise samples to further determine whether the two underwater target radiated noise samples are of the same type of target.

[0011] Furthermore, the multi-channel DEMON spectrum in step S1 is represented as X. i (k), where 0≤i<N, 0≤k<L, N represents the number of channels in the DEMON spectrum processing, and L represents the length of the Fourier transform;

[0012] At each frequency, the spectral lines of channel N reach their maximum values, forming a single-channel demon spectrum. The single-channel demon spectrum is as follows:

[0013]

[0014] Where Y(k) constitutes single-channel data; X i (k) represents the multichannel DEMON spectrum;

[0015] To perform a mean reduction operation on Y(k), the center point of Y(k) is adjusted. The mean reduction operation is as follows:

[0016] For each frequency point, the mean of Y(k) is subtracted from the spectral line, and then compared with 0. The larger value is taken to ensure that the spectral lines of the sample are non-negative. The process of subtracting the mean is expressed as follows:

[0017]

[0018] Normalization is applied to Y′(k), meaning the maximum value of the spectral line cannot exceed 1, within the range [0,1]. The normalization operation is expressed as follows:

[0019]

[0020] Where Z(k) represents the preprocessed sample.

[0021] Furthermore, the underwater target radiated noise with Doppler frequency offset in step S2 is expressed as follows:

[0022] L1(t) = L(ηt)

[0023] Where L(t) is the target radiated noise, η is the scale transformation factor, and η = c / (c+v) ≈ 1 - v / c = 1 - f dv is the relative radial velocity, c is the speed of sound in water, and f is the velocity of the relative radial velocity. d =v / c represents the relative Doppler frequency offset; by changing f d The value of is used to generate underwater target radiation noise with different Doppler frequency offsets;

[0024] Representation of radiated noise of underwater targets at different signal-to-noise ratios:

[0025] L2(t)=L(t)+δN(t)

[0026] Where n(t) is Gaussian white noise; by adjusting the value of the coefficient δ, underwater target radiated noise with different signal-to-noise ratios is generated;

[0027] Underwater target radiation noise with different spectral lines can be generated by changing the number of sinusoidal signals in L(t).

[0028] Furthermore, the method for extracting the sample pairs in step S2 is as follows:

[0029] DEMON processing is applied to the radiated noise of multiple types of underwater targets to obtain multiple samples Z(k), which constitute the training samples. Among the samples of the same type of target, two samples are randomly selected to form a sample pair, called a positive sample pair. Among the samples of any two different types of targets, one sample is randomly selected from each to form a sample pair, called a negative sample pair. The sample extraction and combination process is repeated to obtain multiple positive and negative sample pairs, which are used for training the convolutional neural network model.

[0030] The radiated noise of multiple types of underwater targets is processed by incorporating Doppler frequency offset, signal-to-noise ratio, and interference spectral lines to simulate different working and motion states of the targets. Then, DEMON processing is performed to obtain multiple samples Z(k), which constitute the verification and test samples. From the samples of the same type of target, two samples are randomly selected to form a positive sample pair; from the samples of any two different types of targets, one sample is randomly selected from each to form a negative sample pair. This sample extraction and combination process is repeated to obtain multiple sets of positive and negative sample pairs, which are used for the verification and testing of the convolutional neural network model.

[0031] Label the above sample pairs: label 1 for positive sample pairs and label 0 for negative sample pairs.

[0032] Furthermore, the convolutional neural network model described in step S3 has a symmetrical structure, consisting of two one-dimensional convolutional neural networks sharing weights combined with a fully connected network. The input to the convolutional neural network model is a sample pair consisting of two samples, with each sample containing 2048 data points. The one-dimensional convolutional neural network contains four convolutional layers, four pooling layers, and two fully connected layers. Each convolutional layer has four convolutional kernels with a length of four. Each pooling layer compresses the data volume to half that of the previous pooling layer. The fully connected network outputs 64 data points. The cross-entropy loss is calculated from the two output data of the convolutional neural network model. By minimizing the cross-entropy loss, the convolutional neural network model is trained and optimized. Finally, the convolutional neural network model obtains a similarity function between the two samples, which is used to determine whether the two samples belong to the same target class.

[0033] Compared with existing technologies, the deep learning-based underwater target recognition method with a small sample size described in this invention has the following advantages:

[0034] This invention discloses a method for identifying underwater targets with a small sample size based on deep learning. The innovation lies in the following: when processing the DEMON spectrum of underwater target radiation noise, the maximum value is calculated across multiple channels to obtain a single-channel DEMON spectrum, and then a "mean subtraction" operation is performed. This preserves the characteristics of the DEMON spectrum while reducing the data volume, facilitating network model training. The underwater target radiation noise is processed by incorporating factors such as Doppler frequency offset, signal-to-noise ratio, and interference spectral lines to simulate different working and motion states of the target, generating datasets with different functions. A Siamese network model based on a one-dimensional convolutional neural network is designed to calculate the similarity between two samples and determine whether they belong to the same category, achieving the goal of small sample target identification. When designing the network model, validation sets containing different Doppler frequency offsets, different signal-to-noise ratios, and different numbers of spectral lines are used to evaluate the network performance, improving the network's generalization performance. Attached Figure Description

[0035] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0036] Figure 1 This is a schematic diagram of the processing procedure for the DEMON spectrum of underwater target radiation noise described in an embodiment of the present invention, i.e., the sample generation process;

[0037] Figure 2 This is a schematic diagram illustrating the process of generating sample pairs and datasets for small-sample underwater target radiated noise identification according to an embodiment of the present invention;

[0038] Figure 3This is a schematic diagram of the structure and parameters of the convolutional neural network model for underwater target radiation noise identification as described in an embodiment of the present invention. Detailed Implementation

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0040] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0041] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0042] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0043] like Figure 1 As shown, a method for underwater target recognition with a small sample size based on deep learning includes the following steps:

[0044] (1) Model underwater target radiation noise, give a method for processing underwater target radiation noise DEMON spectrum, and calculate the maximum value between channels of multi-channel DEMON spectrum to obtain single-channel DEMON spectrum, reduce DEMON spectrum data volume, and perform mean subtraction operation on DEMON spectrum as a sample for training and verification of convolutional neural network.

[0045] (2) To address the issue of sample diversity caused by different states of underwater targets, radiation noise is processed to simulate different working and motion states of the targets. Factors such as Doppler frequency offset, signal-to-noise ratio, and interference spectrum are added to generate the training dataset, validation dataset, and test dataset required for the convolutional neural network. These datasets are then combined to generate sample pairs for training, validation, and testing of the Siamese network model.

[0046] (3) Design a Siamese network model based on a one-dimensional convolutional neural network (i.e., the convolutional neural network model mentioned above), train the parameters of the convolutional neural network with the original underwater target radiation noise DEMON spectrum, evaluate the network performance with a validation set containing different Doppler frequency offsets, different signal-to-noise ratios and different numbers of spectral lines, and obtain the convolutional neural network model, which is used to calculate the "similarity" of two underwater target radiation noise samples, and further determine whether the two samples are the same type of target.

[0047] The multi-channel DEMON spectrum described in step (1) is represented as X. i (k), where 0 ≤ i < N, 0 ≤ k < L, N represents the number of channels for DEMON spectrum processing, and L represents the length of the Fourier transform. To reduce the computational cost of the neural network and maximize the preservation of feature information, at each frequency, the spectral lines of the N channels are maximized to form a new DEMON spectrum, which can be described as...

[0048]

[0049] Y(k) constitutes single-channel data, which can minimize the training and recognition time of the neural network.

[0050] Further processing of Y(k) involves adjusting the center point of Y(k), i.e., subtracting the mean of the entire Y(k) from the spectral line at each frequency point, and then comparing it with 0, taking the larger value to ensure that the spectral lines of the sample are non-negative. This process can be represented as:

[0051]

[0052] Normalization is performed on Y′(k), meaning the maximum value of the spectral line cannot exceed 1, within the range [0,1]. This operation is expressed as:

[0053]

[0054] Z(k) represents the preprocessed sample.

[0055] The radiated noise with Doppler frequency offset mentioned in step (2) can be expressed as:

[0056] L1(t) = L(ηt)

[0057] Where L(t) is the target radiated noise, η is the scale transformation factor, and η = c / (c+v) ≈ 1 - v / c = 1 - f d v is the relative radial velocity, c is the speed of sound in water, and f is the velocity of the relative radial velocity. d =v / c represents the relative Doppler frequency offset. By changing f d The value of is used to generate a series of target radiated noises with different Doppler frequency offsets.

[0058] Radiated noise at different signal-to-noise ratios can be expressed as:

[0059] L2(t)=L(t)+δN(t)

[0060] Where n(t) is Gaussian white noise. By adjusting the value of the coefficient δ, a series of target radiated noises with different signal-to-noise ratios are generated.

[0061] Different types of interference-induced radiated noise can be generated by changing the amount of sinusoidal signal in L(t), thus creating a series of target radiated noises with different interference spectra.

[0062] The method for extracting sample pairs in step (2) is as follows.

[0063] DEMON processing is applied to the radiated noise of various underwater target types, resulting in multiple samples Z(k), which constitute the training samples. Among samples of the same type of target, two samples are randomly selected to form a sample pair, called a positive sample pair; among samples of any two different types of targets, one sample is randomly selected from each to form a sample pair, called a negative sample pair. This sample extraction and combination process is repeated to obtain a sufficient number of positive and negative sample pairs for training the network model.

[0064] Multiple types of underwater target radiated noise were processed by incorporating factors such as Doppler frequency offset, signal-to-noise ratio, and interference spectral lines to simulate different target operating and motion states. Then, DEMON processing was performed, resulting in multiple samples Z(k), which constituted the validation and test samples. From samples of the same type of target, two samples were randomly selected to form a positive sample pair; from samples of any two different types of targets, one sample was randomly selected from each to form a negative sample pair. This sample extraction and combination process was repeated to obtain a sufficient number of positive and negative sample pairs for network model validation and testing.

[0065] Label all sample pairs: positive sample pairs are labeled with 1, and negative sample pairs are labeled with 0.

[0066] The Siamese network model based on a one-dimensional convolutional neural network described in step (3) has a symmetrical structure, consisting of two one-dimensional convolutional neural networks sharing weights combined with a fully connected network. The input to the network model is a pair of samples, each with 2048 data points. The one-dimensional convolutional neural network contains 4 convolutional layers, 4 pooling layers, and 2 fully connected layers. Each convolutional layer has 4 kernels with a length of 4, and each pooling layer compresses the data volume to half of its original size. The fully connected network outputs 64 data points. The cross-entropy loss is calculated from the two output data streams. By minimizing the cross-entropy loss, the network is trained and optimized. Finally, the network model can obtain a similarity function between the two samples to determine whether the two samples belong to the same target class.

[0067] Example 1

[0068] The specific embodiments of the present invention will be described in detail below.

[0069] 1. Method for processing DEMON spectrum of radiated noise from underwater targets

[0070] The power spectrum of underwater target radiated noise is a hybrid spectrum composed of line spectra and continuous spectra. This invention primarily identifies the line spectrum, which reflects the presence of certain periodically stable vibration sources within the target, consisting of a series of frequencies f1, f2, ..., f... M It consists of discrete frequency components, and its mathematical model is:

[0071]

[0072] In the formula, f i p is the line spectrum frequency of the i-th discrete frequency point. i Let φ be the line spectral intensity at the i-th discrete frequency point. i The phase is random, and ω is the frequency of the carrier signal.

[0073] The signal model of the actual received underwater target radiated noise signal S(t) is as follows:

[0074] S(t)=L(t)+N(t) (2)

[0075] In the formula, L(t) represents the time-domain signal of the line spectrum, and N(t) represents the marine environmental noise, which is mainly composed of Gaussian white noise.

[0076] When processing radiated noise signals, the received broadband signal is typically demodulated to obtain line spectrum information, a process known as DEMON analysis. The demodulated low-frequency time-domain signal is called the envelope signal, and its power spectrum is the DEMON spectrum. The processing procedure is as follows: Figure 1As shown. Before the DEMON analysis, the received radiated noise signal is first bandpass filtered to obtain the radiated noise signal within the required frequency range. For ease of analysis, consider any single-frequency modulation signal sinΩt; the modulation signal model is then:

[0077] L′(t)=A(1+msinΩt)·cosωt (3)

[0078] In the formula, A represents the amplitude of the signal; m represents the modulation index, 0 < m < 1; ω is the frequency of the carrier signal, and Ω represents the frequency of the modulating signal. Absolute value detection is performed on the signal:

[0079]

[0080] As can be seen from the above equation, the signal |L′(t)| after absolute value detection mainly consists of DC components and harmonic components. Therefore, the signal envelope can be obtained through filtering.

[0081] First, bandpass filter is applied to |L′(t)| to remove the DC and high-frequency components, resulting in the low-frequency signal component. Then, the discrete signal corresponding to the component is downsampled to obtain the time-series signal x(n). A discrete Fourier transform is performed on x(n), and the modulus of the transform result is taken to obtain the DEMON spectrum X(k), where 0≤n≤L-1, 0≤k≤L-1, and L is the length of the Fourier transform.

[0082] The radiated noise is processed by multiple bandpass filters. Taking an example with 6 filters and a Discrete Fourier Transform length of 2048, the explanation is as follows: The DEMON spectrum of the 6 channels can be represented as X. i (k), 0≤i≤5, 0≤k≤2047. To reduce the computational cost of the neural network and preserve feature information to the maximum extent, the spectral lines of the 6 channels are maximized at each frequency to form a new DEMON spectrum, which can be described as...

[0083]

[0084] Y(k) constitutes single-channel data, which can minimize the training and recognition time of the neural network.

[0085] Further processing of Y(k) involves adjusting the center point of Y(k), i.e., subtracting the mean of the entire Y(k) from the spectral line at each frequency point, and then comparing it with 0, taking the larger value to ensure that the spectral lines of the sample are non-negative. This process can be represented as:

[0086]

[0087] Normalization is performed on Y′(k), meaning the maximum value of the spectral line cannot exceed 1, within the range [0,1]. This operation is expressed as:

[0088]

[0089] Where Z(k) represents the preprocessed sample.

[0090] 2. Methods for generating datasets with different functions

[0091] The operational and motion states of a target lead to diverse samples for the same target, while the collected samples may only contain a limited number of states, making it difficult to establish a complete sample library for the same target and causing difficulties in target identification. This invention processes underwater target radiated noise by incorporating factors such as Doppler frequency shift, signal-to-noise ratio, and interference spectral lines to simulate different operational and motion states of the target, generating datasets with different functions. The process of generating the datasets is as follows: Figure 2 As shown.

[0092] Due to the relative motion between the underwater target and the sonar platform, the underwater target radiated noise received by the sonar will exhibit a certain Doppler frequency shift. The signal received by the sonar can be expressed as...

[0093] L1(t)=L(ηt) (8)

[0094] Where L(t) is the target radiated noise, η is the scale transformation factor, and η = c / (c+v) ≈ 1 - v / c = 1 - f d v is the relative radial velocity, c is the speed of sound in water, and f is the velocity of the relative radial velocity. d =v / c represents the relative Doppler frequency offset. By changing f d The value of f is used to generate a series of target radiated noises with different Doppler frequency offsets. These are then processed using the DEMON spectrum to generate samples for the validation and test sets. For example, setting f... d The range is -0.02 to 0.02, with an interval of 0.001, which can generate 41 different Doppler frequency offset validation set samples.

[0095] Due to factors such as marine environmental noise and the distance between the sonar and the target, the signal-to-noise ratio (SNR) of the target radiated noise received by sonar varies considerably. The signal received by sonar can be expressed as...

[0096] L2(t)=L(t)+δN(t) (9)

[0097] Where n(t) is Gaussian white noise. By adjusting the value of the coefficient δ, a series of target radiated noises with different signal-to-noise ratios are generated. Through DEMON spectrum processing, samples for the validation and test sets are generated. For example, by setting the value of δ so that the signal-to-noise ratio varies from 1.0 to 4.0 with an interval of 0.2, 16 validation set samples with different signal-to-noise ratios can be generated.

[0098] Due to environmental or human interference, redundant spectral lines may appear in the radiated noise received by sonar, or some spectral lines may weaken or even disappear. During simulation, according to formula (1), changing the number M of the sinusoidal signal in L(t) can generate a series of target radiated noises with different interference spectral lines. Through DEMON spectral processing, samples for the validation set and test set can be generated. For example, setting the number of spectral lines to M-3 to M+4 can generate 8 samples of different interferences.

[0099] The target radiated noise with different parameters such as Doppler frequency offset, signal-to-noise ratio and interference spectrum is processed by the improved DEMON method to obtain the preprocessed sample Z(k).

[0100] When designing a Siamese network model for underwater target recognition with a small sample size, the input data for the network model consists of two samples. The network calculates a similarity function between the two samples to further determine whether they belong to the same type of target. Therefore, multiple samples need to be further processed to obtain the sample pairs required by the Siamese network.

[0101] like Figure 2 As shown, DEMON processing is applied to the radiated noise of multiple underwater target types, resulting in multiple samples Z(k), which constitute the training samples. Among samples of the same type of target, two samples are randomly selected to form a sample pair, called a positive sample pair; among samples of any two different types of targets, one sample is randomly selected from each to form a sample pair, called a negative sample pair. This sample extraction and combination process is repeated to obtain a sufficient number of positive and negative sample pairs for training the network model.

[0102] Multiple types of underwater target radiated noise were processed by incorporating factors such as Doppler frequency offset, signal-to-noise ratio, and interference spectral lines to simulate different target operating and motion states. Then, DEMON processing was performed, resulting in multiple samples Z(k), which constituted the validation and test samples. From samples of the same type of target, two samples were randomly selected to form a positive sample pair; from samples of any two different types of targets, one sample was randomly selected from each to form a negative sample pair. This sample extraction and combination process was repeated to obtain a sufficient number of positive and negative sample pairs for network model validation and testing.

[0103] Label all sample pairs: positive sample pairs are labeled with 1, and negative sample pairs are labeled with 0.

[0104] 3. Design of a Siamese Network Model Based on a One-Dimensional Convolutional Neural Network

[0105] The structure and parameters of the Siamese network model based on a one-dimensional convolutional neural network for identifying radiated noise from underwater targets with a small sample size, as proposed in this invention, are as follows: Figure 3 As shown.

[0106] The Siamese network model has a symmetrical structure, consisting of two one-dimensional convolutional neural networks sharing weights combined with a fully connected network. The input to the network model is a pair of samples, each with 2048 data points. The one-dimensional convolutional neural network contains four convolutional layers, four pooling layers, and two fully connected layers. Each convolutional layer has four kernels of length 4, and each pooling layer compresses the data volume to half its original size. The fully connected network outputs 64 data points. Cross-entropy loss is calculated between the two output streams. The network is trained and optimized by minimizing this cross-entropy loss. Finally, the network model obtains a similarity function between the two samples, determining whether they belong to the same target class.

[0107] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for underwater target recognition with a small sample size based on deep learning, characterized in that: Includes the following steps: S1. Model underwater target radiated noise, give a method for processing underwater target radiated noise DEMON spectrum, find the maximum value between channels of multi-channel DEMON spectrum to obtain single-channel DEMON spectrum, and perform mean subtraction operation on single-channel DEMON spectrum as a sample for training and validation of convolutional neural network; S2. Process the underwater target radiated noise, simulate different working and motion states of the underwater target radiated noise, add Doppler frequency offset, signal-to-noise ratio, and interference spectrum lines, generate the training dataset, validation dataset, and test dataset required for the convolutional neural network, and combine them to generate sample pairs, which are used for training, validation, and testing of the convolutional neural network model. S3. Design a convolutional neural network model: The parameters of the convolutional neural network were trained using the original DEMON spectrum of underwater target radiated noise. The network performance was evaluated using a validation set with different Doppler frequency offsets, different signal-to-noise ratios, and different numbers of spectral lines to obtain the convolutional neural network model. The convolutional neural network model was used to calculate the similarity function of two underwater target radiated noise samples to further determine whether the two underwater target radiated noise samples are of the same type. The underwater target radiated noise with Doppler frequency offset in step S2 is expressed as follows: ； in, For target radiated noise, The scaling factor. , The relative radial velocity, The speed of sound in water, Indicates the relative Doppler frequency shift; by changing The value of is used to generate underwater target radiation noise with different Doppler frequency offsets; Representation of radiated noise of underwater targets at different signal-to-noise ratios: ； in, Gaussian white noise; adjustment factor The value of is used to generate underwater target radiated noise with different signal-to-noise ratios; Underwater target radiated noise with different spectral lines, by changing The number of sinusoidal signals generates underwater target radiation noise with different interference spectra; The method for extracting the sample pairs in step S2 is as follows: DEMON processing was performed on the radiated noise of multiple types of underwater targets, resulting in multiple samples. The samples are used to form training samples. Among the samples of the same type of target, two samples are randomly selected to form a sample pair, called a positive sample pair. Among the samples of any two different types of targets, one sample is randomly selected from each to form a sample pair, called a negative sample pair. The sample extraction and combination process is repeated to obtain multiple sets of positive and negative sample pairs, which are used for training the convolutional neural network model. The radiated noise of multiple types of underwater targets was processed by incorporating Doppler frequency shift, signal-to-noise ratio, and interference spectral lines to simulate different working and motion states of the targets. Then, DEMON processing was performed to obtain multiple samples. The samples are used to form verification and test samples. Two samples are randomly selected from the samples of the same type of target to form a positive sample pair. One sample is randomly selected from each of the samples of any two different types of targets to form a negative sample pair. The sample extraction and combination process is repeated to obtain multiple positive and negative sample pairs for verification and testing of the convolutional neural network model. Label the above sample pairs: label 1 for positive sample pairs and label 0 for negative sample pairs. The convolutional neural network model in step S3 has a symmetrical structure, consisting of two one-dimensional convolutional neural networks sharing weights combined with a fully connected network. The input to the convolutional neural network model is a pair of samples, each with 2048 data points. The one-dimensional convolutional neural network contains four convolutional layers, four pooling layers, and two fully connected layers. Each convolutional layer has four kernels with a length of four. Each pooling layer compresses the data to half the size of the previous pooling layer. The fully connected network outputs 64 data points. The two output data streams of the convolutional neural network model are used to calculate the cross-entropy loss. By minimizing the cross-entropy loss, the convolutional neural network model is trained and optimized. Finally, the convolutional neural network model obtains a similarity function between the two samples, which is used to determine whether the two samples belong to the same target class.

2. The method for underwater target recognition based on deep learning with a small sample size according to claim 1, characterized in that: The multi-channel DEMON spectrum in step S1 is represented as follows: ,in , , This indicates the number of channels processed by the DEMON spectrum. Indicates the length of the Fourier transform; At each frequency, The spectral lines of each channel are taken at their maximum values ​​to form a single-channel demon spectrum. The single-channel demon spectrum is as follows: ； in, This constitutes single-channel data; Multichannel DEMON spectrum; right Perform a mean subtraction operation to adjust. The mean subtraction operation at the center point is as follows: Spectral lines at each frequency point minus The mean is calculated, then compared with 0, and the larger value is taken to ensure that the spectral lines of the sample are non-negative. The process of subtracting the mean is expressed as follows: ； right Normalization is performed, meaning the maximum value of the spectral line cannot exceed 1, within the range [0,1]. The normalization operation is expressed as follows: ； in, This is the sample after preprocessing.

3. An electronic device, comprising a processor and a memory communicatively connected to the processor and used for storing processor-executable instructions, characterized in that: The processor is used to execute the deep learning-based underwater target recognition method with a small sample size as described in any one of claims 1-2.

4. A server, characterized in that: It includes at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor to cause the at least one processor to perform a deep learning-based method for identifying underwater targets with a small sample size as described in any one of claims 1-2.

5. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the deep learning-based method for identifying underwater targets with a small sample size as described in any one of claims 1-2.

Images