A method and device for identifying underwater targets

Abstract

The present invention discloses an underwater target recognition method and device. The method comprises: obtaining an underwater target radiation noise data set, comprising N pieces of underwater target radiation noise data information, where N is an integer; extracting characteristic parameters of any piece of underwater target radiation noise data information to obtain characteristic information of the underwater target radiation noise, including radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information; the underwater target radiation noise characteristic information constitutes an underwater target radiation noise characteristic information set; dividing the underwater target radiation noise characteristic information set to obtain a training set and a test set; using the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model; using the training underwater target recognition model to process the test set to obtain an underwater target recognition result.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition, and in particular to a method and a device for underwater target recognition. Background Art

[0002] Judging the target category based on the information reflected by the underwater target radiation noise signal is a common underwater target identification method. The characteristics of underwater acoustic target radiation noise signal are that the sound sources are numerous and concentrated, the spectrum composition is diverse, and the radiation noise intensity is large. Radiated noise is usually composed of continuously distributed broadband noise and narrowband line spectrum components distributed on several discrete frequencies. Low-frequency line spectrum is mainly generated by periodic motions such as mechanical rotation and propeller blade rotation. It has high intensity, and the center frequency is located at low frequency and the long-distance propagation loss is small. It is an important information source for passive detection, tracking and identification of targets. Affected by the fluctuation of the radiation energy of the sound source itself and the multipath effect, the amplitude of the line spectrum has certain fluctuations. At the same time, affected by the slight changes of the line spectrum source rotation speed and the Doppler frequency shift caused by the change of relative motion parameters, the frequency of the line spectrum will also have certain changes. Therefore, the stability of amplitude and frequency is also an important feature of the line spectrum. Mastering the amplitude stability characteristics of the target radiation noise line spectrum can provide a basis for the selection and optimization of the amplitude threshold in the line spectrum feature extraction method, and understanding the frequency stability characteristics has important guiding significance for the selection and optimization of the frequency threshold in the line spectrum feature extraction method.

[0003] The composition, generation mechanism and transmission path of radiation noise are relatively complex, which makes it present nonlinear and non-stationary characteristics. A single time domain or frequency domain feature cannot comprehensively represent the effective information of the signal in the time and frequency scale, and cannot obtain the overall change of the signal in the time and frequency domain. Therefore, it is necessary to study the method of multi-feature fusion to obtain more representative features.

[0004] After extracting the signal features, the target signal is identified by constructing a classifier. The classifier is trained to learn the signal feature vector to obtain the discrimination rule for identifying the target. Under this discrimination rule, the original data set of the signal is mapped to the feature space, and the identified target objects are classified into one category and the non-target objects are classified into another category by using statistical methods. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method and device for underwater target recognition, which can extract the radiation noise line spectrum, modulation spectrum, continuous spectrum and auditory spectrum characteristics of the target, combine the support vector machine classifier method to carry out decision layer fusion, and realize online recognition research of the target type.

[0006] In order to solve the above technical problems, the first aspect of the embodiment of the present invention discloses a method for identifying an underwater target, the method comprising:

[0007] S1, obtaining an underwater target radiation noise data set, wherein the underwater target radiation noise data set includes N pieces of underwater target radiation noise data information, where N is an integer;

[0008] S2, extracting characteristic parameters of any piece of underwater target radiation noise data information to obtain characteristic information of underwater target radiation noise;

[0009] The underwater target radiation noise characteristic information includes radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information;

[0010] The underwater target radiation noise characteristic information constitutes an underwater target radiation noise characteristic information set;

[0011] S3, dividing the characteristic information set of the underwater target radiation noise to obtain a training set and a test set;

[0012] S4, using the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model;

[0013] S5, using the training underwater target recognition model to process the test set to obtain an underwater target recognition result.

[0014] As an optional implementation manner, in the first aspect of the embodiment of the present invention, the method for calculating the radiation noise line spectrum characteristic information includes:

[0015] Perform power spectrum analysis on any piece of underwater target radiation noise data information to obtain a power spectrum diagram;

[0016] Performing filtering on the power spectrum to obtain a flattened power spectrum;

[0017] Processing the flattened power spectrum using a preset amplitude threshold and a preset frequency threshold to obtain a target radiation noise line spectrum;

[0018] Processing the target radiation noise line spectrum to obtain radiation noise line spectrum feature information;

[0019] The radiation noise line spectrum characteristic information includes the mean value of the line spectrum amplitude, the relative standard deviation of the line spectrum amplitude, the mean value of the line spectrum center frequency, and the relative standard deviation of the line spectrum center frequency.

[0020] As an optional implementation manner, in the first aspect of the embodiment of the present invention, the method for calculating the modulation spectrum characteristic information includes:

[0021] Using the envelope time domain model, any data information of the radiation noise of the underwater target is processed to obtain the time domain envelope information;

[0022] Using the envelope frequency domain model, any data information of the radiation noise of the underwater target is processed to obtain the frequency domain envelope information;

[0023] The time domain envelope information and the frequency domain envelope information constitute modulation spectrum feature information.

[0024] As an optional implementation manner, in the first aspect of the embodiment of the present invention, the method for calculating the continuous spectrum feature information includes:

[0025] Process any piece of underwater target radiation noise data information to obtain a time-frequency spectrum diagram;

[0026] Selecting four observation planes on the time-frequency spectrum diagram;

[0027] On the four observation planes, obtaining point sets of their contour lines;

[0028] The effective frequency range of the time-frequency spectrum is divided into M segments, and the number of intersections between each segment and the point set of the contour line is counted according to the frequency to obtain feature vector information;

[0029] Processing the characteristic vector information to obtain peak frequency, rising rate and decay rate;

[0030] The peak frequency, rise rate and decay rate constitute continuous spectrum characteristic information.

[0031] As an optional implementation manner, in a first aspect of the embodiments of the present invention, the method for calculating the auditory spectrum feature information includes:

[0032] Preprocessing any piece of underwater target radiation noise data information to obtain preprocessed radiation noise data information;

[0033] Dividing the preprocessed radiation noise data information into frequency bands to obtain critical frequency band radiation noise data information;

[0034] Filtering the critical frequency band radiation noise data information to obtain filtered data information;

[0035] The filter data information is subjected to cepstrum calculation to obtain auditory spectrum feature information.

[0036] As an optional implementation, in the first aspect of the embodiment of the present invention, the using of the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model includes:

[0037] S41, setting the search range of the penalty parameter C and the RBF kernel parameter σ of the preset underwater target recognition model;

[0038] S42, generating a two-dimensional grid along different growth directions of the penalty parameter C and the RBF kernel parameter σ within the search range with a preset fixed search step size, wherein the nodes in the two-dimensional grid correspond to each set of C and σ values;

[0039] S43, calculating the multi-fold cross validation accuracy of each group of C and σ under the training set;

[0040] S44, taking the group of C and σ with the highest multi-fold cross-validation accuracy as the best parameter model parameters;

[0041] S45, using the optimal parameter model parameters to train a preset underwater target recognition model to obtain a trained underwater target recognition model.

[0042] As an optional implementation manner, in the first aspect of the embodiment of the present invention, the using the training water target recognition model to process the test set to obtain the water target recognition result includes:

[0043] S51, preset n training water target recognition models;

[0044] S52, using n training underwater target recognition models to process the test set to obtain n test results;

[0045] S53, using the Bayesian reasoning method to perform decision fusion on the n test results to obtain an underwater target recognition result.

[0046] The second aspect of the present invention discloses an underwater target recognition device, the device comprising:

[0047] A data acquisition module is used to acquire an underwater target radiation noise data set, wherein the underwater target radiation noise data set includes N pieces of underwater target radiation noise data information, where N is an integer;

[0048] A feature extraction module is used to extract feature parameters from any piece of underwater target radiation noise data information to obtain feature information of underwater target radiation noise;

[0049] The underwater target radiation noise characteristic information includes radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information;

[0050] The underwater target radiation noise characteristic information constitutes an underwater target radiation noise characteristic information set;

[0051] A data partitioning module is used to partition the underwater target radiation noise characteristic information set to obtain a training set and a test set;

[0052] A training module, used to train a preset underwater target recognition model using the training set to obtain a training underwater target recognition model;

[0053] The recognition module is used to process the test set using the training underwater target recognition model to obtain an underwater target recognition result.

[0054] As an optional implementation manner, in a second aspect of the embodiment of the present invention, the method for calculating the radiation noise line spectrum characteristic information includes:

[0055] Perform power spectrum analysis on any piece of underwater target radiation noise data information to obtain a power spectrum diagram;

[0056] Performing filtering on the power spectrum to obtain a flattened power spectrum;

[0057] Processing the flattened power spectrum using a preset amplitude threshold and a preset frequency threshold to obtain a target radiation noise line spectrum;

[0058] Processing the target radiation noise line spectrum to obtain radiation noise line spectrum feature information;

[0059] The radiation noise line spectrum characteristic information includes the mean value of the line spectrum amplitude, the relative standard deviation of the line spectrum amplitude, the mean value of the line spectrum center frequency, and the relative standard deviation of the line spectrum center frequency.

[0060] As an optional implementation manner, in the second aspect of the embodiment of the present invention, the method for calculating the modulation spectrum characteristic information includes:

[0061] Using the envelope time domain model, any data information of the radiation noise of the underwater target is processed to obtain the time domain envelope information;

[0062] Using the envelope frequency domain model, any data information of the radiation noise of the underwater target is processed to obtain the frequency domain envelope information;

[0063] The time domain envelope information and the frequency domain envelope information constitute modulation spectrum feature information.

[0064] As an optional implementation manner, in the second aspect of the embodiment of the present invention, the method for calculating the continuous spectrum feature information includes:

[0065] Process any piece of underwater target radiation noise data information to obtain a time-frequency spectrum diagram;

[0066] Selecting four observation planes on the time-frequency spectrum diagram;

[0067] On the four observation planes, obtaining point sets of their contour lines;

[0068] The effective frequency range of the time-frequency spectrum is divided into M segments, and the number of intersections between each segment and the point set of the contour line is counted according to the frequency to obtain feature vector information;

[0069] Processing the characteristic vector information to obtain peak frequency, rising rate and decay rate;

[0070] The peak frequency, rise rate and decay rate constitute continuous spectrum characteristic information.

[0071] As an optional implementation manner, in a second aspect of the embodiment of the present invention, the method for calculating the auditory spectrum feature information includes:

[0072] Preprocessing any piece of underwater target radiation noise data information to obtain preprocessed radiation noise data information;

[0073] Dividing the preprocessed radiation noise data information into frequency bands to obtain critical frequency band radiation noise data information;

[0074] Filtering the critical frequency band radiation noise data information to obtain filtered data information;

[0075] The filter data information is subjected to cepstrum calculation to obtain auditory spectrum feature information.

[0076] As an optional implementation, in the second aspect of the embodiment of the present invention, the using of the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model includes:

[0077] S41, setting the search range of the penalty parameter C and the RBF kernel parameter σ of the preset underwater target recognition model;

[0078] S42, generating a two-dimensional grid along different growth directions of the penalty parameter C and the RBF kernel parameter σ within the search range with a preset fixed search step size, wherein the nodes in the two-dimensional grid correspond to each set of C and σ values;

[0079] S43, calculating the multi-fold cross validation accuracy of each group of C and σ under the training set;

[0080] S44, taking the group of C and σ with the highest multi-fold cross-validation accuracy as the best parameter model parameters;

[0081] S45, using the optimal parameter model parameters to train a preset underwater target recognition model to obtain a trained underwater target recognition model.

[0082] As an optional implementation, in the second aspect of the embodiment of the present invention, the using the training water target recognition model to process the test set to obtain the water target recognition result includes:

[0083] S51, preset n training water target recognition models;

[0084] S52, using n training underwater target recognition models to process the test set to obtain n test results;

[0085] S53, using the Bayesian reasoning method to perform decision fusion on the n test results to obtain an underwater target recognition result.

[0086] The third aspect of the present invention discloses another underwater target recognition device, the device comprising:

[0087] A memory storing executable program code;

[0088] a processor coupled to the memory;

[0089] The processor calls the executable program code stored in the memory to execute part or all of the steps in the underwater target recognition method disclosed in the first aspect of the embodiment of the present invention.

[0090] The fourth aspect of the present invention discloses a computer storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute some or all of the steps in the underwater target recognition method disclosed in the first aspect of the embodiment of the present invention.

[0091] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0092] In the embodiment of the present invention, by extracting the radiation noise line spectrum, modulation spectrum, continuous spectrum and auditory spectrum features of the underwater target, combined with the support vector machine classifier method to carry out decision-level fusion, online identification research of the target type is realized. The decision-level fusion in target recognition is based on the preliminary completion of target classification for each feature, imitating human thinking to make combined judgments with the help of decision-making fusion methods, and integrating the different signal attributes highlighted by different features at the decision-making layer, thereby improving the accuracy of system discrimination. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0094] Figure 1 It is a flow chart of a method for identifying underwater targets disclosed in an embodiment of the present invention;

[0095] Figure 2 It is a principle block diagram of a target radiation noise line spectrum feature extraction method and stability analysis disclosed in an embodiment of the present invention;

[0096] Figure 3 is a schematic diagram of line spectrum analysis parameters disclosed in an embodiment of the present invention;

[0097] Figure 4 It is a schematic diagram of a method for extracting features of a modulation spectrum of radiation noise disclosed in an embodiment of the present invention;

[0098] Figure 5 is a flow chart of target propeller modulation feature extraction based on model matching disclosed in an embodiment of the present invention;

[0099] Figure 6 A flowchart of auditory feature extraction based on an auditory peripheral model disclosed in an embodiment of the present invention;

[0100] Figure 7 A flowchart of the target radiated noise auditory feature extraction disclosed in an embodiment of the present invention;

[0101] Figure 8 A flowchart of the target radiated noise auditory feature extraction disclosed in an embodiment of the present invention;

[0102] Fig. 9 Contour map of SVM parameter selection results based on the grid search algorithm disclosed in an embodiment of the present invention;

[0103] Fig.10 Flow chart of the target fusion recognition method disclosed in the embodiment of the present invention;

[0104] Fig.11 A schematic diagram of the structure of an underwater target recognition device disclosed in an embodiment of the present invention;

[0105] Fig.12 A schematic structural diagram of another underwater target recognition device disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0106] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0107] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or equipment that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or equipment.

[0108] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0109] The present invention discloses a method and device for identifying underwater targets, which includes: obtaining an underwater target radiation noise data set, including N underwater target radiation noise data information, where N is an integer; extracting characteristic parameters of any underwater target radiation noise data information to obtain underwater target radiation noise characteristic information, including radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information; the underwater target radiation noise characteristic information constitutes an underwater target radiation noise characteristic information set; the underwater target radiation noise characteristic information set is divided to obtain a training set and a test set; the training set is used to train a preset underwater target recognition model to obtain a training underwater target recognition model; the training underwater target recognition model is used to process the test set to obtain an underwater target recognition result. The following are detailed descriptions.

[0110] Embodiment 1

[0111] See also Figure 1 , Figure 1 : is a flow chart of a method for identifying underwater targets disclosed in an embodiment of the present invention. Figure 1 The underwater target recognition method described is applied to the field of underwater signal recognition technology, and the embodiments of the present invention do not limit this. Figure 1 As shown, the underwater target recognition method may include the following operations:

[0112] S1, obtaining an underwater target radiation noise data set, wherein the underwater target radiation noise data set includes N pieces of underwater target radiation noise data information, where N is an integer;

[0113] S2, extracting characteristic parameters of any piece of underwater target radiation noise data information to obtain characteristic information of underwater target radiation noise;

[0114] The underwater target radiation noise characteristic information includes radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information;

[0115] The underwater target radiation noise characteristic information constitutes an underwater target radiation noise characteristic information set;

[0116] S3, dividing the characteristic information set of the underwater target radiation noise to obtain a training set and a test set;

[0117] The ratio of the training set to the test set is 8:1;

[0118] S4, using the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model;

[0119] S5, using the training underwater target recognition model to process the test set to obtain an underwater target recognition result.

[0120] Optionally, the method for calculating the radiation noise line spectrum characteristic information includes:

[0121] Perform power spectrum analysis on any piece of underwater target radiation noise data information to obtain a power spectrum diagram;

[0122] Performing filtering on the power spectrum to obtain a flattened power spectrum;

[0123] Processing the flattened power spectrum using a preset amplitude threshold and a preset frequency threshold to obtain a target radiation noise line spectrum;

[0124] Processing the target radiation noise line spectrum to obtain radiation noise line spectrum feature information;

[0125] The radiation noise line spectrum characteristic information includes the mean value of the line spectrum amplitude, the relative standard deviation of the line spectrum amplitude, the mean value of the line spectrum center frequency, and the relative standard deviation of the line spectrum center frequency.

[0126] Optionally, the method for calculating the modulation spectrum characteristic information includes:

[0127] Using the envelope time domain model, any data information of the radiation noise of the underwater target is processed to obtain the time domain envelope information;

[0128] Using the envelope frequency domain model, any data information of the radiation noise of the underwater target is processed to obtain the frequency domain envelope information;

[0129] The time domain envelope information and the frequency domain envelope information constitute modulation spectrum feature information.

[0130] Optionally, the method for calculating the continuous spectrum feature information includes:

[0131] Process any piece of underwater target radiation noise data information to obtain a time-frequency spectrum diagram;

[0132] Selecting four observation planes on the time-frequency spectrum diagram;

[0133] On the four observation planes, obtaining point sets of their contour lines;

[0134] The effective frequency range of the time-frequency spectrum is divided into M segments, and the number of intersections between each segment and the point set of the contour line is counted according to the frequency to obtain feature vector information;

[0135] Processing the characteristic vector information to obtain peak frequency, rising rate and decay rate;

[0136] The peak frequency, rise rate and decay rate constitute continuous spectrum characteristic information.

[0137] Optionally, the method for calculating the auditory spectrum feature information includes:

[0138] Preprocessing any piece of underwater target radiation noise data information to obtain preprocessed radiation noise data information;

[0139] Dividing the preprocessed radiation noise data information into frequency bands to obtain critical frequency band radiation noise data information;

[0140] Filtering the critical frequency band radiation noise data information to obtain filtered data information;

[0141] The filter data information is subjected to cepstrum calculation to obtain auditory spectrum feature information.

[0142] Optionally, the using the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model includes:

[0143] S41, setting the search range of the penalty parameter C and the RBF kernel parameter σ of the preset underwater target recognition model;

[0144] S42, generating a two-dimensional grid along different growth directions of the penalty parameter C and the RBF kernel parameter σ within the search range with a preset fixed search step size, wherein the nodes in the two-dimensional grid correspond to each set of C and σ values;

[0145] S43, calculating the multi-fold cross validation accuracy of each group of C and σ under the training set;

[0146] S44, taking the group of C and σ with the highest multi-fold cross-validation accuracy as the best parameter model parameters;

[0147] S45, using the optimal parameter model parameters to train a preset underwater target recognition model to obtain a trained underwater target recognition model.

[0148] Optionally, the using the training underwater target recognition model to process the test set to obtain an underwater target recognition result includes:

[0149] S51, preset n training water target recognition models;

[0150] S52, using n training underwater target recognition models to process the test set to obtain n test results;

[0151] S53, using the Bayesian reasoning method to perform decision fusion on the n test results to obtain an underwater target recognition result.

[0152] Embodiment 2

[0153] 1. Radiated noise line spectrum feature extraction method

[0154] Radiated noise is usually composed of continuously distributed broadband noise and narrowband line spectrum components distributed on several discrete frequencies. Low-frequency line spectrum is mainly generated by periodic motions such as mechanical rotation and propeller blade rotation. It has high intensity, and the center frequency is at low frequency and has low long-range propagation loss. It is an important information source for passive detection, tracking and identification of targets.

[0155] The present invention establishes a method for extracting the line spectrum characteristics of target radiation noise and carries out stability analysis, such as Figure 2 As shown. Based on the measured data of the target radiation noise characteristics, the power spectrum analysis is performed using the classic spectrum estimation method based on Fourier transform to obtain the power spectrum. The trend items in the power spectrum are filtered out using a two-way α filter or a low-pass filter to obtain a flattened spectrum. The amplitude threshold is set to remove the spectrum components that are less than the amplitude threshold. The frequency threshold is set to remove the spectrum components that are wider than the frequency threshold to obtain the target radiation noise line spectrum. For the line spectrum, analysis is carried out from the two perspectives of amplitude and frequency. Figure 3 The frequency f0 and relative amplitude ΔL are marked on the simulated line spectrum. p An intuitive representation of .

[0156] Affected by the fluctuation of the radiation energy of the sound source itself and the multipath effect, the amplitude of the line spectrum has certain fluctuations. At the same time, affected by the slight changes in the rotation speed of the line spectrum source and the Doppler frequency shift caused by the change of relative motion parameters, the frequency of the line spectrum will also have certain changes. Therefore, the stability of amplitude and frequency is also an important feature of the line spectrum. Mastering the amplitude stability characteristics of the target radiation noise line spectrum can provide a basis for the selection and optimization of the amplitude threshold in the line spectrum feature extraction method, and understanding the frequency stability characteristics has important guiding significance for the selection and optimization of the frequency threshold in the line spectrum feature extraction method.

[0157] The present invention establishes a method for analyzing the amplitude and frequency stability of line spectrum, and uses the relative standard deviation of line spectrum amplitude and frequency to describe the stability of line spectrum. The mean value of line spectrum amplitude is:

[0158]

[0159] Where A i is the amplitude of the line spectrum of the i-th signal. The relative standard deviation of the line spectrum amplitude is:

[0160]

[0161] Where S A is the standard deviation of the line spectrum amplitude.

[0162] The mean of the center frequencies of the line spectra is:

[0163]

[0164] Where f i is the center frequency of the line spectrum of the i-th signal. The relative standard deviation of the line spectrum center frequency is

[0165]

[0166] Where S f is the standard deviation of the center frequency of the line spectrum.

[0167] 2. Target radiated noise modulation spectrum feature extraction method

[0168] Based on the study of the propeller blade envelope formation mechanism, the propeller radiated noise envelope waveform characteristics are obtained through the analysis and statistics of a large number of measured samples, and the propeller noise envelope model is established through the long-term accumulation of prior knowledge such as existing target data and target types, propeller attribute parameters, and the mining of the modulation characteristics of target radiated noise. The model uses parameters such as the number of propellers, the number of blades, the time difference between propellers, the amplitude of blade radiated sound pulses, the pulse shape function, and the propeller rotation period to describe the propeller modulation envelope characteristics. On the basis of the envelope model, the multi-propeller radiated noise envelope model, the modulation line spectrum characteristics, and the envelope waveform structure characteristics are combined, and the attribute parameter characteristics of the target propeller, such as the number of propellers and the number of blades, are extracted through the model matching method. Figure 4 It is a schematic diagram of the method for extracting the characteristics of the radiated noise modulation spectrum disclosed in the embodiment of the present invention. The modulation envelope waveform under multiple rotation cycles has basically equal intervals; each "package" has a basically consistent shape; the height of each "package" has a certain randomness; for the modulation of the blade frequency, these "packages" also meet certain regular grouping structures and other characteristics. Therefore, the noise modulation envelope can be treated as a pulsed random process with the same shape, the same repetition period, random amplitude and multi-propeller superposition in a grouped structure, and a propeller envelope model can be established based on this.

[0169] When propeller blades rotate in an uneven wake field, they will excite acoustic pulses with amplitude modulation properties. These acoustic pulses are related to the blade modulation degree, wake field distribution, etc. An N-blade propeller will radiate N acoustic pulses for each rotation (each blade radiates one acoustic pulse in one cycle, which can be represented by an envelope). Therefore, with the continuous rotation of the N-blade propeller, a series of acoustic pulses will be radiated. These acoustic pulses have the following properties: within a certain period of time, the wake field distribution can be considered stable, so the random acoustic pulses radiated by the same blade in different propeller rotation cycles are similar. Moreover, since the physical properties of the blades on the same propeller are also similar, the acoustic pulses radiated by different blades are also similar. However, due to physical differences in the manufacturing process and blade corrosion caused by long-term work, the modulation degrees of different blades are different. Therefore, the acoustic pulses radiated by different blades in the same cycle will show a strong and weak arrangement, showing the properties of a blade pulse amplitude group structure.

[0170] By studying the basic characteristics of the target propeller noise envelope waveform, using time domain waveform analysis and statistical analysis techniques combined with frequency domain analysis, the statistical characteristics of the envelope waveform structure and amplitude of different target blades are studied, and a target propeller radiated noise envelope time domain model with blade grouping structure, bounded random pulse shape, random amplitude, and pulse random process of the same repetition period that conforms to the envelope characteristics is established.

[0171] ① Time domain model of propeller cavitation noise envelope

[0172]

[0173] T m is the rotation period of the mth propeller, M represents the number of propellers, R represents the number of blades, and m represents the mth propeller. and It is the pulse shape function and pulse amplitude function of the blade radiation excited when the rth blade of the mth propeller rotates one circle in the non-uniform wake field.

[0174] ② Envelope frequency domain model of ship propeller cavitation noise

[0175]

[0176] In the formula, is the lth order harmonic of the modulation line spectrum of the noise radiated by M blades R. (mm′) (l) is the cross-product term of the m-th and m'-th propellers, reflecting the interaction between the two propellers. It is the modulation line spectrum of the noise radiated by the mth propeller alone, that is, the modulation line spectrum model of single propeller noise. It is the cross-multiplication sum of M propellers, which is related to the phase difference between each propeller and is the basis for identifying the number of propellers of multi-propeller ships.

[0177] The target propeller radiated noise envelope model, modulation line spectrum characteristics and envelope waveform structure characteristics are combined to establish a matching model. The optimal propeller radiated noise envelope model is found through optimization theory, and the best match between the measured target time domain envelope signal and the model envelope is sought. The target propeller modulation feature extraction method based on model matching is studied. In the optimization process of the matching model, the modulation envelope waveform is used as the matching parameter of the nonlinear programming model, and the genetic algorithm is used to find the optimal value in the search space so that the matching model can reach the optimal parameter value and obtain the attribute parameter information such as the number of blades and the number of blades of the target propeller. Figure 5 This is a flow chart of target propeller modulation feature extraction based on model matching disclosed in an embodiment of the present invention. The present invention combines the target propeller radiation noise envelope model, modulation line spectrum characteristics and envelope waveform structure characteristics to establish a matching model, finds the optimal propeller radiation noise envelope model through optimization theory, seeks the best match between the measured target time domain envelope signal and the model envelope, and studies the target propeller modulation feature extraction method based on model matching. In the optimization process of the matching model, the modulation envelope waveform and the like are used as matching parameters of the nonlinear programming model, and the genetic algorithm is used to optimize in the search space so that the matching model reaches the optimal parameter value, and the attribute parameter information such as the number of propellers and the number of blades of the target propeller is obtained.

[0178] 3. Research on the auditory characteristics and continuous spectrum feature extraction method of target radiated noise

[0179] (1) Method for extracting auditory features of radiated noise

[0180] Well-trained listeners can hear weak target radiation noise signals from the complex ocean background and can accurately identify the category of the target. This is the advantage of human hearing. The time-frequency-amplitude characteristics of the sound signal, especially the perception of special timbre, are unmatched by ordinary signal processing algorithms. Without considering the influence of the complex noise environment of the ocean, from the perspective of auditory perception, the difference in the radiation noise of different categories of targets is obvious. It can be imagined that if we can accurately understand and grasp the human auditory processing mechanism, and establish a model that can be implemented on a computer, extract the important features perceived by the human auditory system, and apply it to underwater target recognition, it will definitely be able to greatly improve the performance of the recognition system.

[0181] The present invention integrates the Gammatone auditory filter that can simulate the frequency decomposition characteristics of the basilar membrane and the Meddis model that simulates the transduction characteristics of the inner hair cells to represent the processing mechanism of the human cochlea, and modifies the model parameters according to the feature extraction effect of the measured data, and establishes the Gammatone-Meddis auditory perception representation model. The underwater target radiation noise signal is processed by the model to obtain the auditory nerve firing rate, and combined with the listener's hearing judgment type, the auditory perception features that the listener is interested in and can reflect the essence of the target are extracted. Figure 6 A flowchart of auditory feature extraction based on an auditory peripheral model disclosed in an embodiment of the present invention;

[0182] The specific steps of auditory feature extraction are as follows: Figure 7 As shown in the figure, the target radiation noise signal in the water is firstly preprocessed by frame segmentation and mean removal, and the signal frequency is divided according to the critical frequency band of the human ear (Bark scale) within the spectrum range of the signal. According to the divided frequencies, several Gammatone-Meddis auditory filter groups are set, and the radiation noise signal is convolved by several auditory filter groups with different frequency bands. Then, the output of the auditory filter group is processed by energy and cepstrum, so that a variety of auditory features can be extracted from different angles, which is also an indirect description of the loudness and timbre of the radiation noise of different types of targets. Figure 8 A flow chart of the extraction of auditory features of target radiated noise disclosed in an embodiment of the present invention.

[0183] (2) Target radiated noise continuous spectrum feature extraction method

[0184] Time-frequency analysis describes the change of the spectral content of a signal over time. It is a local change that can reflect the energy strength of each frequency at each moment. After applying the short-time Fourier transform method to obtain the time-frequency spectrum, the task is how to extract effective classification information that can express the characteristics of the target continuous spectrum from the spectrum to form a feature vector. The technical route of the continuous spectrum feature extraction method adopted by the present invention is: first, in the time-frequency spectrum, four observation planes are selected, and then the point set of their contour lines is taken; then the effective frequency range of the analyzed signal is divided into M segments, and the number of intersections of each segment is counted according to the frequency, that is, an M-dimensional feature vector is obtained.

[0185] The continuous spectrum feature vector graphs of the three types of targets have good separability and can be combined with classifiers to achieve classification and identification of target types.

[0186] The present invention also studies a method for extracting a continuous spectrum of radiation noise based on power spectrum analysis. This method extracts its "trend term" based on the classic power spectrum analysis of the radiation noise signal based on Fourier transform, removes the influence of line spectrum, and extracts a relatively gentle continuous spectrum. According to a large amount of data statistics, the continuous spectrum of the target radiation noise usually has a peak value. Below the peak frequency, the continuous spectrum rises with a certain positive slope, and above the peak frequency, the continuous spectrum decreases with a certain negative slope, as shown in the figure below. According to the characteristics of the target continuous spectrum, typical features such as the continuous spectrum, peak frequency, rise rate and attenuation rate can be extracted.

[0187] 4. Research on online target data identification technology based on the radiation noise characteristics of underwater targets

[0188] The present invention adopts the existing support vector machine classifier as the local decision maker, applies a grid search algorithm to a sample set of underwater target noise characteristics, realizes parameter selection of the support vector machine classifier, and improves the classification accuracy and generalization ability of the support vector machine.

[0189] In the application of SVM, there is still a problem that the optimal model parameters need to be selected according to different feature samples so that SVM can achieve the optimal classification effect. From the research of Vapnik et al., it is known that although different kernel functions have a certain influence on SVM performance, the kernel parameters and penalty parameters in the kernel function are the key factors affecting SVM performance. Therefore, under the premise of determining the use of RBF kernel parameters, the present invention establishes an optimization selection problem with kernel parameters σ and penalty parameter C as variables, applies a grid search algorithm to solve the optimization problem and determine the optimal values ​​of SVM parameters, so that the performance of the SVM classifier is stably improved.

[0190] The main process of parameter selection for support vector machine classifiers implemented by applying grid search algorithm is as follows: first determine the search range of penalty parameter C and RBF kernel parameter σ, then generate a two-dimensional grid along different growth directions of the two parameters with a fixed search step size. The nodes in the grid are parameter pairs consisting of all possible penalty factor parameters and kernel parameters within a given range, corresponding to each set of C and σ values ​​on the grid. Calculate the multi-fold cross-validation accuracy of each set of C and σ in the training set, and finally draw the accuracy corresponding to each set of C and σ values ​​with contour lines to obtain a contour map. Finally, take the set of C and σ that has the highest classification accuracy of the training set as the optimal parameter, as shown in Fig. 9 shown.

[0191] The process of using the support vector machine classifier to identify different types of targets is as follows: first, the training set and test set are extracted from the feature set of the underwater target radiation noise, and then normalization preprocessing is performed. The training set is used to select the best parameters C and σ, and then the SVM is trained with the selected best parameters. The trained model is then used to predict the test set to obtain the classification accuracy.

[0192] Based on the support vector machine classifier, the present invention conducts research on the decision-layer fusion method. The decision-layer fusion in target recognition is to combine and judge each discrimination result based on the preliminary target classification of each feature by imitating human thinking with certain rules or algorithms, that is, to fuse and decide on the different signal attributes highlighted by different features at the decision layer to improve the accuracy of system discrimination. Fig.10 As shown in Figure 1, it is an important model for target fusion recognition in practical work. Fig.10 The local decision maker adopts the support vector machine classifier, and the decision layer fusion mainly studies the decision fusion methods such as Bayesian reasoning and voting method. According to the characteristics of actual underwater acoustic signals, the two algorithms are combined, and a joint method of voting and Bayesian reasoning is proposed. The feature template, support vector machine classifier and decision fusion algorithm are combined to realize the classification and recognition of multiple types of targets.

[0193] The standard of the decision fusion method proposed in the present invention is: if voting has a clear resolution to the problem, the result of the voting method is accepted first, otherwise the fusion is performed according to the Bayesian fusion strategy. The support vector machine classification results of the three features of the auditory feature product feature, continuous spectrum feature and high-order spectrum feature of the passive noise signal are used for multi-feature comprehensive recognition.

[0194] Optionally, the radiated noise line spectrum feature information, modulation spectrum feature information, continuous spectrum feature information and auditory spectrum feature information may be fused to obtain fused feature information.

[0195] Optionally, the method for fusing feature information is to first fuse the radiated noise line spectrum feature information and the modulation spectrum feature information to obtain a first fused feature; fuse the continuous spectrum feature information and the auditory spectrum feature information to obtain a second fused feature; fuse the first fused feature and the second fused feature to obtain fused feature information. The specific method for fusing the two feature parameters is:

[0196] Input feature vector sample set X = [x1, x2, ..., x n ] and feature sample set Y = [y1,y2,…,y m ].

[0197] 1. Initialization

[0198] Find the autocovariance matrix ∑ of sample X 11 , the covariance matrix of Y∑ 22 and the cross-covariance matrix of X and Y∑ 12 and∑ 21 ; is the between-class matrix, is the intra-class matrix;

[0199] 2. Construct the Lagrange equation according to the optimization goal

[0200]

[0201] Where, Q1 = ∑ 11 , Q2=∑ 22 , β, λ, and θ are coefficient constants obtained through multiple experiments and are not limited in the present invention.

[0202] The optimization goal is to minimize Make

[0203] 3. To W x and W y Find partial derivatives and transform the optimization problem into the solution of eigenvalues

[0204] 4. Apply the eigendecomposition method to obtain the matrix composed of the eigenvectors corresponding to the non-zero vectors and the transformation matrix W of X x .

[0205] 5. According to Get the transformation matrix W of Y y .

[0206] 6. Substitute the transformation matrix into the formula In the above example, we get the final fusion feature Σ k .

[0207] Fusion feature information

[0208] Another optional feature fusion method is to concatenate the radiated noise line spectrum feature information, modulation spectrum feature information, continuous spectrum feature information and auditory spectrum feature information. Assuming there are n samples in total and the feature dimension of each sample is m, the input matrix of the fusion method is:

[0209] X={x 1 ,x 2 ,...,x n} m×n

[0210] In the above formula, each x represents a sample and is an m-dimensional column vector.

[0211] Step 1: Data centralization.

[0212]

[0213] According to the above formula, we first find the mean of each sample, which is an m*1-dimensional vector. Then, we can complete the data centering by subtracting the mean vector from each column of the original feature matrix. The purpose of data centering is to enhance the orthogonality between the original features.

[0214] Step 2: Based on the data center, calculate the covariance matrix of the feature matrix

[0215] Step 3: Perform eigenvalue decomposition on the obtained covariance matrix, sort the m eigenvalues ​​in descending order, and extract the eigenvectors corresponding to the first k eigenvalues. These eigenvectors are the projection matrix. The K value is selected based on the first k principal components λ i The cumulative contribution rate is defined as When the cumulative contribution rate reaches a certain proportion, the original feature can be replaced.

[0216] Step 3: According to the projection matrix obtained in the previous step, the original features are projected to obtain the reduced-dimensional features. Assume that the feature matrix after projection is Y = {y 1 ,y 2 ,...,y n} k×n , the first k eigenvectors calculated based on the first k eigenvalues ​​extracted are u1,u2,...,u k .but:

[0217] y 1 =u1X=u 11 x 1 +u 12 x 2 +...+u 1m x m

[0218] y 2 =u2X=u 21 x 1 +u 22 x 2 +...+u 2m x m ...

[0219] y k =u k X=u k1 x 1 +u k2 x 2 +...+u km x m

[0220] Optionally, the least squares support vector machine (LS-SVM) is used as the recognition model, and the expression of its optimization solution process is:

[0221]

[0222] sty i = <w,φ(x i )>+b+e i ,

[0223] i=1,2,…n,γ>0.

[0224] Among them, x i ∈R d is the sample data, y i {+1,-1};e i are the corresponding category labels and discrimination errors respectively, i=1,2,…,Ntr.

[0225] The corresponding dual problem is:

[0226]

[0227] The optimal KKT condition is:

[0228]

[0229] After eliminating w and e by elimination method, it can be further expressed as:

[0230]

[0231] Where y=[y1,y2,…,y n ] T ; 1 v =[1,1,…,1] T ; α=[α1,α2,…αn ] T .

[0232] Optional, such as Fig.10 The radiated noise line spectrum feature information, modulation spectrum feature information, continuous spectrum feature information and auditory spectrum feature information are input into a local decision maker based on feature template matching, and decision fusion is performed to obtain the final decision result. The process is as follows:

[0233] Divide the feature matrix (any one of the radiated noise line spectrum feature information, modulation spectrum feature information, continuous spectrum feature information and auditory spectrum feature information) into a training set M1 and a test set M2;

[0234] Randomly select a certain proportion of samples M from the training set M1 with replacement 1i (i represents the i-th decision tree) as the input sample of a decision tree;

[0235] Construct n decision trees through the classification and regression tree method, and take the classification result with the highest probability of all decision trees as the result of the random forest classifier;

[0236] When n changes from 1 to 200, the best accuracy of the classifier is the number of corresponding decision trees, thus obtaining a local decision maker based on feature template matching;

[0237] Optionally, the feature matrix (any one of the radiated noise line spectrum feature information, modulation spectrum feature information, continuous spectrum feature information and auditory spectrum feature information) may be preprocessed to improve the accuracy of target recognition. The specific method is as follows:

[0238] 1. Multi-space representation learning, given two input networks G1 and G2, G1 and G2 use VGG19 network, VGG19 network is a prior art, other network models can also be used, the present invention does not limit. In order to simultaneously learn the regular structure and hierarchical structure features in the network space, this method convolves the original network in different geometric spaces to obtain the neighbor information of the nodes in the network, and obtains the Euclidean space embedding representation and hyperbolic space embedding representation of each node.

[0239] 2. The Euclidean space embedding representation and the hyperbolic space embedding representation are fused to obtain fused features. The present invention does not limit the specific fusion method.

[0240] Euclidean space network embedding: Each network can be represented by two matrices: the adjacency matrix A and the node feature matrix X (if the node has no features, it can be the identity matrix), where each row x in X i Represents the characteristics of the node. First, normalize the adjacency matrix A:

[0241]

[0242] Where I is the identity matrix, is the diagonal matrix, D i,i =∑ j A ij .

[0243] In order to obtain the network structure representation in Euclidean space, it is necessary to perform graph convolution on the input feature matrix, which follows the following feedforward pass:

[0244]

[0245] in is the Euclidean space parameter matrix of the specified layer; σ(·) is a nonlinear function, such as ReLU(·)=max(0,·); is the node embedding matrix of the lth layer, the input layer Z 0 =X; d is the embedding dimension of each node. This feed-forward transfer enables each node to obtain the information of its neighbor nodes through the normalized adjacency matrix, thereby learning the network structure.

[0246] Hyperbolic space network embedding: The initial node features belong to the Euclidean space, which is first mapped to the hyperbolic space. First, use express The origin in (0,x i ) can be considered as A point in is mapped to hyperbolic space as follows

[0247] H 0 =exp o (X)

[0248] For a given network, a feed-forward pass of a (l+1)-layer graph convolution to generate a node embedding matrix in hyperbolic space is as follows:

[0249]

[0250] in is the hyperbolic space parameter matrix of the lth layer, is a hyperbolic transformation, AGG(·) is a neighborhood aggregation operation in a hyperbolic space, It is a hyperbolic nonlinear activation function.

[0251] Embodiment 3

[0252] See also Fig.11 , Fig.11 : is a schematic diagram of the structure of an underwater target recognition device disclosed in an embodiment of the present invention. Fig.11The underwater target recognition device described is applied to the field of underwater signal recognition technology, and the embodiments of the present invention are not limited thereto. Fig.11 As shown, the underwater target identification device may include the following operations:

[0253] S301, a data acquisition module, configured to acquire an underwater target radiation noise data set, wherein the underwater target radiation noise data set includes N pieces of underwater target radiation noise data information, where N is an integer;

[0254] S302, a feature extraction module is used to extract feature parameters from any piece of underwater target radiation noise data information to obtain feature information of the underwater target radiation noise;

[0255] The underwater target radiation noise characteristic information includes radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information;

[0256] The underwater target radiation noise characteristic information constitutes an underwater target radiation noise characteristic information set;

[0257] S303, a data division module, used to divide the characteristic information set of the underwater target radiation noise to obtain a training set and a test set;

[0258] S304, a training module, used to train a preset underwater target recognition model using the training set to obtain a training underwater target recognition model;

[0259] S305, a recognition module is used to process the test set using the training underwater target recognition model to obtain an underwater target recognition result.

[0260] Embodiment 4

[0261] See also Fig.12 , Fig.12 is a schematic diagram of the structure of another underwater target recognition device disclosed in an embodiment of the present invention. Fig.12 The underwater target recognition device described is applied to the field of underwater signal recognition technology, and the embodiments of the present invention are not limited thereto. Fig.12 As shown, the underwater target identification device may include the following operations:

[0262] A memory 401 storing executable program codes;

[0263] a processor 402 coupled to the memory 401;

[0264] The processor 402 calls the executable program code stored in the memory 401 to execute the steps in the underwater target recognition method described in the first embodiment or the second embodiment.

[0265] Embodiment 4

[0266] An embodiment of the present invention discloses a computer-readable storage medium storing a computer program for electronic data exchange, wherein the computer program enables a computer to execute the steps of the underwater target recognition method described in the first or second embodiment.

[0267] The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, i.e., they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0268] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution can be essentially or partly contributed to the prior art in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0269] Finally, it should be noted that the underwater target recognition method and device disclosed in the embodiments of the present invention only disclose the preferred embodiments of the present invention, which are only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying underwater targets, characterized in that: The method comprises: S1, obtaining an underwater target radiation noise data set, wherein the underwater target radiation noise data set includes N pieces of underwater target radiation noise data information, where N is an integer; S2, processing any piece of underwater target radiation noise data information to obtain underwater target radiation noise characteristic information; The underwater target radiation noise characteristic information includes radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information; S21, the method for calculating the radiation noise line spectrum characteristic information includes: Perform power spectrum analysis on any piece of underwater target radiation noise data information to obtain a power spectrum diagram; Performing filtering on the power spectrum to obtain a flattened power spectrum; Processing the flattened power spectrum using a preset amplitude threshold and a preset frequency threshold to obtain a target radiation noise line spectrum; Processing the target radiation noise line spectrum to obtain radiation noise line spectrum feature information; The radiation noise line spectrum characteristic information includes the mean value of the line spectrum amplitude, the relative standard deviation of the line spectrum amplitude, the mean value of the line spectrum center frequency and the relative standard deviation of the line spectrum center frequency; S22, the method for calculating the modulation spectrum characteristic information includes: Using the envelope time domain model, any data information of the radiation noise of the underwater target is processed to obtain the time domain envelope information; The envelope time domain model is: T m is the rotation period of the mth propeller, M represents the number of propellers, R represents the number of blades, and m represents the mth propeller. and is the pulse shape function and pulse amplitude function of the blade radiation excited when the rth blade of the mth propeller rotates one circle in the non-uniform wake field; Using the envelope frequency domain model, any data information of the radiation noise of the underwater target is processed to obtain the frequency domain envelope information; The envelope frequency domain model is: Where P R M (l) is the lth order harmonic of the modulation line spectrum of the radiation noise of M R blades; P (mm′) (l) is the cross-product term of the m-th and m'-th propellers, reflecting the interaction between the two propellers; is the modulation line spectrum of the noise radiated by the mth propeller alone, i.e., the modulation line spectrum model of the noise of a single propeller; It is the cross-product sum of M propellers, which is related to the phase difference between each propeller and is the basis for identifying the number of propellers in a multi-propeller ship. The time domain envelope information and the frequency domain envelope information constitute modulation spectrum feature information; S23, the method for calculating the continuous spectrum characteristic information includes: Process any piece of underwater target radiation noise data information to obtain a time-frequency spectrum diagram; Selecting four observation planes on the time-frequency spectrum diagram; On the four observation planes, obtaining point sets of their contour lines; The effective frequency range of the time-frequency spectrum is divided into M segments, and the number of intersections between each segment and the point set of the contour line is counted according to the frequency to obtain feature vector information; Processing the characteristic vector information to obtain peak frequency, rising rate and decay rate; The peak frequency, rise rate and decay rate constitute continuous spectrum characteristic information S24, the method for calculating the auditory spectrum feature information includes: Preprocessing any piece of underwater target radiation noise data information to obtain preprocessed radiation noise data information; The pre-processing comprises: Multi-space representation learning: Given two input networks G1 and G2, G1 and G2 use the VGG19 network, and convolve the original network in different geometric spaces to obtain the neighbor information of the nodes in the network, and obtain the Euclidean space embedding representation and hyperbolic space embedding representation of each node; The Euclidean space embedding representation and the hyperbolic space embedding representation are fused to obtain fused features; the fused features are preprocessed radiation noise data information; Dividing the preprocessed radiation noise data information into frequency bands to obtain critical frequency band radiation noise data information; Filtering the critical frequency band radiation noise data information to obtain filtered data information; Performing cepstrum calculation on the filtered data information to obtain auditory spectrum feature information; S25, fusing the radiated noise line spectrum feature information, the modulation spectrum feature information, the continuous spectrum feature information and the auditory spectrum feature information to obtain fused feature information, including: The method for fusing feature information is to first fuse the radiated noise line spectrum feature information and the modulation spectrum feature information to obtain a first fusion feature; fuse the continuous spectrum feature information and the auditory spectrum feature information to obtain a second fusion feature; fuse the first fusion feature and the second fusion feature to obtain fusion feature information; The specific fusion method of the two feature parameters is: Input feature vector sample set X = [x1, x2, ..., x n ] and feature sample set Y = [y1,y2,…,y m ]; initialization Find the autocovariance matrix v of sample X 11 , the covariance matrix of Y∑ 22 and the cross-covariance matrix of X and Y∑ 12 and∑ 21 ; is the between-class matrix, is the intra-class matrix; According to the optimization goal, construct the Lagrangian equation Where, Q1 = ∑ 11 , Q2=∑ 22 , β, λ, θ are coefficient constants; The optimization goal is to minimize Make To W x and W y Find partial derivatives and transform the optimization problem into solving eigenvalues; Apply the eigendecomposition method to obtain the matrix composed of the corresponding eigenvectors of the non-zero vectors and the transformation matrix W of X x ; according to Get the transformation matrix W of Y y ; Substituting the transformation matrix into the formula In the above example, we get the final fusion feature Σ k ; The fused feature information constitutes a feature information set of radiation noise of underwater targets; S3, dividing the characteristic information set of the underwater target radiation noise to obtain a training set and a test set; S4, using the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model; S5, using the training underwater target recognition model, processing the test set of characteristic information of radiation noise of the underwater target to be recognized, and obtaining the underwater target recognition result.

2. The underwater target recognition method according to claim 1, characterized in that: The method of using the training set to train a preset underwater target recognition model to obtain a training underwater target recognition model includes: S41, setting the search range of the penalty parameter C and the RBF kernel parameter σ of the preset underwater target recognition model; S42, generating a two-dimensional grid along different growth directions of the penalty parameter C and the RBF kernel parameter σ within the search range with a preset fixed search step size, wherein the nodes in the two-dimensional grid correspond to each set of C and σ values; S43, calculating the multi-fold cross validation accuracy of each group of C and σ under the training set; S44, taking the set of C and σ with the highest multi-fold cross-validation accuracy as the best model parameters; S45, using the optimal model parameters to train a preset underwater target recognition model to obtain a trained underwater target recognition model.

3. The underwater target recognition method according to claim 1, characterized in that: The method of using the training underwater target recognition model to process the test set to obtain an underwater target recognition result includes: S51, preset n training water target recognition models; S52, using n training water target recognition models to process the test set to obtain n test results; S53, using the Bayesian reasoning method to perform decision fusion on the n test results to obtain an underwater target recognition result.

4. An underwater target recognition device, characterized in that: The device comprises: A data acquisition module is used to acquire an underwater target radiation noise data set, wherein the underwater target radiation noise data set includes N pieces of underwater target radiation noise data information, where N is an integer; A feature extraction module is used to process any piece of underwater target radiation noise data information to obtain the underwater target radiation noise feature information; The underwater target radiation noise characteristic information includes radiation noise line spectrum characteristic information, modulation spectrum characteristic information, continuous spectrum characteristic information and auditory spectrum characteristic information; S21, the method for calculating the radiation noise line spectrum characteristic information includes: Perform power spectrum analysis on any piece of underwater target radiation noise data information to obtain a power spectrum diagram; Performing filtering on the power spectrum to obtain a flattened power spectrum; Processing the flattened power spectrum using a preset amplitude threshold and a preset frequency threshold to obtain a target radiation noise line spectrum; Processing the target radiation noise line spectrum to obtain radiation noise line spectrum feature information; The radiation noise line spectrum characteristic information includes the mean value of the line spectrum amplitude, the relative standard deviation of the line spectrum amplitude, the mean value of the line spectrum center frequency and the relative standard deviation of the line spectrum center frequency; S22, the method for calculating the modulation spectrum characteristic information includes: Using the envelope time domain model, any data information of the radiation noise of the underwater target is processed to obtain the time domain envelope information; The envelope time domain model is: T m is the rotation period of the mth propeller, M represents the number of propellers, R represents the number of blades, and m represents the mth propeller. and is the pulse shape function and pulse amplitude function of the blade radiation excited when the rth blade of the mth propeller rotates one circle in the non-uniform wake field; Using the envelope frequency domain model, any data information of the radiation noise of the underwater target is processed to obtain the frequency domain envelope information; The envelope frequency domain model is: In the formula, is the lth order harmonic of the modulation line spectrum of the radiation noise of M R blades; P (mm′) (l) is the cross-product term of the m-th and m'-th propellers, reflecting the interaction between the two propellers; is the modulation line spectrum of the noise radiated by the mth propeller alone, i.e., the modulation line spectrum model of the noise of a single propeller; It is the cross-product sum of M propellers, which is related to the phase difference between each propeller and is the basis for identifying the number of propellers in a multi-propeller ship. The time domain envelope information and the frequency domain envelope information constitute modulation spectrum feature information; S23, the method for calculating the continuous spectrum characteristic information includes: Process any piece of underwater target radiation noise data information to obtain a time-frequency spectrum diagram; Selecting four observation planes on the time-frequency spectrum diagram; On the four observation planes, obtaining point sets of their contour lines; The effective frequency range of the time-frequency spectrum is divided into M segments, and the number of intersections between each segment and the point set of the contour line is counted according to the frequency to obtain feature vector information; Processing the characteristic vector information to obtain peak frequency, rising rate and decay rate; The peak frequency, rise rate and decay rate constitute continuous spectrum characteristic information S24, the method for calculating the auditory spectrum feature information includes: Preprocessing any piece of underwater target radiation noise data information to obtain preprocessed radiation noise data information; The pre-processing comprises: Multi-space representation learning: Given two input networks G1 and G2, G1 and G2 use the VGG19 network, and convolve the original network in different geometric spaces to obtain the neighbor information of the nodes in the network, and obtain the Euclidean space embedding representation and hyperbolic space embedding representation of each node; The Euclidean space embedding representation and the hyperbolic space embedding representation are fused to obtain fused features; the fused features are preprocessed radiation noise data information; Dividing the preprocessed radiation noise data information into frequency bands to obtain critical frequency band radiation noise data information; Filtering the critical frequency band radiation noise data information to obtain filtered data information; Performing cepstrum calculation on the filtered data information to obtain auditory spectrum feature information; S25, fusing the radiated noise line spectrum feature information, the modulation spectrum feature information, the continuous spectrum feature information and the auditory spectrum feature information to obtain fused feature information, including: The method for fusing feature information is to first fuse the radiated noise line spectrum feature information and the modulation spectrum feature information to obtain a first fusion feature; fuse the continuous spectrum feature information and the auditory spectrum feature information to obtain a second fusion feature; fuse the first fusion feature and the second fusion feature to obtain fusion feature information; The specific fusion method of the two feature parameters is: Input feature vector sample set X = [x1, x2, ..., x n ] and feature sample set Y = [y1,y2,…,y m ]; initialization Find the autocovariance matrix ∑ of sample X 11 , the covariance matrix of Y∑ 22 and the cross-covariance matrix of X and Y∑ 12 and∑ 21 ; is the between-class matrix, is the intra-class matrix; According to the optimization goal, construct the Lagrangian equation Where, Q1 = ∑ 11 , Q2=∑ 22 , β, λ, θ are coefficient constants; The optimization goal is to minimize Make To W x and W y Find partial derivatives and transform the optimization problem into solving eigenvalues; Apply the eigendecomposition method to obtain the matrix composed of the corresponding eigenvectors of the non-zero vectors and the transformation matrix W of X x ; according to Get the transformation matrix W of Y y ; Substituting the transformation matrix into the formula In the above example, we get the final fusion feature Σ k ; The fused feature information constitutes a feature information set of radiation noise of underwater targets; A data partitioning module, used for partitioning the characteristic information set of the underwater target radiation noise to obtain a training set and a test set; A training module, used to train a preset underwater target recognition model using the training set to obtain a training underwater target recognition model; The recognition module is used to process the test set of radiation noise characteristic information of the underwater target to be recognized by using the training underwater target recognition model to obtain the underwater target recognition result.

5. An underwater target recognition device, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the underwater target recognition method as described in any one of claims 1-3.

6. A computer storable medium, characterized in that: The computer storage medium stores computer instructions, which, when called, are used to execute the underwater target recognition method as described in any one of claims 1 to 3.

Images