CN113780242A - Cross-scene underwater sound target classification method based on model transfer learning - Google Patents

Cross-scene underwater sound target classification method based on model transfer learning Download PDF

Info

Publication number
CN113780242A
CN113780242A CN202111160013.1A CN202111160013A CN113780242A CN 113780242 A CN113780242 A CN 113780242A CN 202111160013 A CN202111160013 A CN 202111160013A CN 113780242 A CN113780242 A CN 113780242A
Authority
CN
China
Prior art keywords
target
scene
model
cross
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111160013.1A
Other languages
Chinese (zh)
Inventor
王大宇
罗恒光
张博轩
王晓庆
唐立赫
张锦灿
王志欣
李鹏飞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
CETC 54 Research Institute
Original Assignee
CETC 54 Research Institute
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by CETC 54 Research Institute filed Critical CETC 54 Research Institute
Priority to CN202111160013.1A priority Critical patent/CN113780242A/en
Publication of CN113780242A publication Critical patent/CN113780242A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Signal Processing (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)

Abstract

The invention discloses a cross-scene underwater sound target classification method based on model transfer learning, belongs to the technical field of underwater sound target passive reconnaissance, and is suitable for classifying underwater sound target signals received based on an array form. The method takes the target signal characteristics extracted by the time-frequency analysis method as a shared characteristic space sample of cross-scene classification, trains a convolutional neural network, and firstly encodes knowledge in a source data domain to form stable classification capability in model parameters, model prior knowledge and a model framework. And then, moving to a target scene, and fitting class relation knowledge in the model in a new sample and label space based on a knowledge distillation technology to realize classification capability. The method solves the problem of poor generalization capability of the existing underwater sound target classification method in cross-scene application through a model migration learning technology, can quickly adapt to new environment and new task under the condition of small sample and data imbalance, and greatly saves time, energy and resources consumed by large amount of data acquisition of a target domain and model retraining.

Description

Cross-scene underwater sound target classification method based on model transfer learning
Technical Field
The invention belongs to the technical field of underwater sound target passive reconnaissance, and particularly relates to a cross-scene underwater sound target classification method based on model transfer learning, which can be used for classifying underwater sound target signals received based on an array form, judging target types and rapidly adapting to different application scenes to keep good classification capability.
Background
The underwater sound target classification is an information processing technology for analyzing and processing target radiation noise signals received by sonar equipment, extracting target characteristics and judging target types. The commonly used target classification methods mainly include statistical classification, model matching, expert system, deep neural network, and the like. When the existing method is applied in a cross-scene mode, due to the reasons of working environment change, target type difference, small samples and the like, the existing method is weak in generalization capability, a classification model needs to be retrained, and in addition, the data period under the condition of obtaining and labeling a new application scene is long and large in consumption, so that the method is an important challenge in the technical result conversion process at the current stage.
The transfer learning aims to solve the learning problem of small data amount in a target scene, and effectively transfers the model from one scene to another scene. The potential motivation of model-based transfer learning is that transferred knowledge is encoded on model levels such as model parameters, model prior knowledge and architecture, an effective model is verified in a specific scene, and classification relation knowledge is extracted by combining a knowledge distillation technology to perform discrimination and classification, so that the cost of performing cross-scene transfer can be greatly reduced. Therefore, the cross-scene underwater sound target classification and identification method based on model migration is an effective solution. However, such a solution is still lacking in the prior art.
Disclosure of Invention
In view of this, the technical problem to be solved by the present invention is to provide a cross-scene underwater sound target classification method based on model transfer learning, which has the advantages of strong generalization capability, high accuracy and good robustness in cross-scene application.
In order to achieve the purpose, the invention adopts the technical scheme that:
a cross-scene underwater sound target classification method based on model transfer learning comprises the following steps:
(1) filtering and denoising an underwater sound target radiation noise signal received by sonar equipment, filtering out clutter and enhancing the signal to form processed time domain signal data;
(2) performing advanced feature mapping processing on the time domain signal data obtained in the step (1), and extracting a time-frequency spectrogram with inter-class discrimination as cross-scene identification shared space features;
(3) determining hyper-parameters of a convolutional neural network model, marking the class of the time-frequency spectrogram, and training the convolutional neural network model;
(4) introducing a knowledge distillation technology into the convolutional neural network model obtained in the step (3), extracting a target type soft label and training, so that class relation knowledge in the model is fitted to a classification performance target in a new sample and label space;
(5) carrying out cross-scene underwater sound target classification by using the convolutional neural network model trained in the step (4);
and completing the cross-scene underwater sound target classification based on model transfer learning.
Wherein, the concrete mode of the step (1) is as follows:
(101) dividing real number form underwater acoustic target radiation noise signals collected by a matrix into equal-time-length samples;
(102) carrying out beam forming on the sample data by adopting a split beam correlation method, and carrying out spectrum peak detection in all directions to select a target suspected direction;
(103) carrying out time accumulation on the suspected target position obtained in the step (102), and obtaining a target accurate position through empirical threshold judgment;
(104) performing time domain beam forming according to the target accurate position obtained in the step (103), and extracting enhanced time domain signal data;
(105) and (5) performing frequency domain filtering on the time domain signal data obtained in the step (104) through an FIR filter to finish filtering and noise reduction.
Wherein, the concrete mode of step (2) is:
and performing short-time Fourier transform on each frame of time domain signal data, and extracting a time-frequency spectrogram with inter-class discrimination as cross-scene identification shared space characteristics.
The convolutional neural network model in the step (3) comprises 1 input layer, 4 convolutional layers, 4 maximum pooling layers, 2 full-link layers and 1 Softmax layer, and the convolutional layers use a ReLU function as an activation function.
In the step (3), the time-frequency spectrogram is labeled to belong to the category, and the convolutional neural network model is trained in the specific mode:
(301) labeling the time spectrum graph according to the target type recorded in the collection process, and using the labeled time spectrum graph as a training sample of the convolutional neural network model;
(302) dividing all time-frequency spectrogram characteristic data into a training set, a verification set and a test set, and disordering the sequence to ensure that various types of data in each sample set are uniformly distributed;
(303) and training the convolutional neural network model by using the sample set data to obtain the convolutional neural network model with the classification and recognition capability.
The specific method of the step (4) is as follows:
(401) adjusting the temperature parameter of a Softmax function in the convolutional neural network model;
(402) extracting cross-scene classification sharing spatial features of data in a target scene for labeling, and taking the cross-scene classification sharing spatial features as training samples;
(403) inputting the training sample into a convolution layer of a convolution neural network model, and passing a nonlinear activation function through the convolution layer;
(404) sending the output of the nonlinear activation function into a pooling layer, performing feature dimension reduction, and reserving key information;
(405) adjusting corresponding parameters, repeating the steps (403) and (404) for a set number of times, and inputting the result into the full connection layer;
(406) inputting the output result of the full connection layer into a softmax function, and extracting a target type soft tag;
(407) calculating the loss of the target type soft label output in the step (406) and the label value, training the soft label, and forming label space and class relation knowledge mapping under a new scene;
(408) and after the training is finished, adjusting the temperature parameter of the Softmax function to be 1, and deploying and applying.
The beneficial effects obtained by the invention are as follows:
1. the method uses an artificial intelligence-based underwater acoustic target identification method, determines a time-frequency spectrogram with inter-class characterization characteristics as a shared space characteristic of cross-scene classification, learns target characteristics by using a deep neural network, and forms target identification capability in a source data domain.
2. The invention is based on the transfer learning technology of the model, utilizes the knowledge obtained by the source data domain to influence the learning and the performance in the target scene, indicates that the knowledge coded into the model parameters, the model prior knowledge and the model architecture is beneficial to the learning of the target domain model, and carries out knowledge sharing on the model level, thereby improving the generalization capability of the model applied across scenes.
3. The invention introduces knowledge distillation technology, solves the practical application problem of small sample amount or unbalance in a target scene, enables the method to be quickly adapted to a new environment, a new target or a new task, can achieve good classification performance by only needing a small amount of target domain samples, and provides a technical approach for the conversion and application of research results.
Drawings
FIG. 1 is a schematic diagram of an embodiment of the present invention;
FIG. 2 is a flow chart of a filtering and denoising process in an embodiment of the present invention;
FIG. 3 is a flow chart of feature extraction in an embodiment of the present invention;
FIG. 4 is a schematic diagram of a convolutional neural network structure for source domain knowledge learning in an embodiment of the present invention;
fig. 5 is a schematic structural diagram of a convolutional neural network model based on a model migration technique in an embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further explained in detail with reference to the attached drawings.
A cross-scene underwater sound target classification method based on model transfer learning is used for preprocessing target noise signals acquired by an array by using a signal processing method to filter interference when extracting features. Fig. 1 is a schematic diagram of implementation of cross-scene training, in which source scene target features are first extracted as training samples, and then a Convolutional Neural Network (CNN) is used to learn sample feature knowledge to form stable classification capability. And then, transferring the model to a target scene, and introducing a knowledge distillation technology to quickly realize the underwater sound target classification capability in the target scene by utilizing the structure and parameters of the source scene model. The method specifically comprises the following steps:
(1) filtering and denoising an underwater sound target radiation noise signal received by sonar equipment, filtering out clutter and enhancing the signal to form processed time domain signal data; the method comprises the following specific steps:
(101) dividing real number form underwater acoustic target radiation noise signals collected by a matrix into equal-time-length samples;
(102) carrying out beam forming on the sample data by adopting a split beam correlation method, and carrying out spectrum peak detection in all directions to select a target suspected direction;
(103) carrying out time accumulation on the suspected target position obtained in the step (102), and obtaining a target accurate position through empirical threshold judgment;
(104) performing time domain beam forming according to the target accurate position obtained in the step (103), and extracting enhanced time domain signal data;
(105) and (5) performing frequency domain filtering on the time domain signal data obtained in the step (104) through an FIR filter to finish filtering and noise reduction.
(2) Performing domain transformation on the time domain signal data in the step (1), and extracting a time frequency spectrogram (LOFAR) feature sample with inter-class discrimination as a cross-scene recognition shared space feature; the method specifically comprises the following steps:
and performing short-time Fourier transform on each frame of time domain signal data, and extracting a time-frequency spectrogram with inter-class discrimination as cross-scene identification shared space characteristics.
(3) Labeling the class of the time-frequency spectrogram characteristic sample obtained in the step (2), and training a Convolutional Neural Network (CNN) by using the time-frequency characteristic; the training skill of 'dropout' can be used in the network training to prevent the over-fitting phenomenon;
the training process comprises the following steps:
(301) classifying and labeling the time spectrum graph characteristic samples according to the target types recorded during collection to serve as training samples of the convolutional neural network;
(302) dividing all time-frequency spectrogram data into a training set, a verification set and a test set according to the proportion of 3:1:1, and disordering the sequence to ensure that various data in each sample set are uniformly distributed;
(303) constructing a convolutional neural network model according to the characteristics of time-frequency data, wherein the convolutional neural network model comprises 1 input layer, 4 convolutional layers, 4 maximum pooling layers, 2 full-connection layers and 1 Softmax layer, and the convolutional layers use a ReLU function as an activation function;
(304) in the training stage, firstly, trainable parameters and hyper-parameters are initialized, time-frequency data are input into CNN to extract characteristic information, and the characteristic information passes through a nonlinear activation function;
(305) sending the output of the activation function to a pooling layer, performing feature dimension reduction, and reserving key information;
(306) adjusting corresponding parameters, repeating the steps (304) and (305) for three times, and inputting the result into a full connection layer to enable the characteristics to be mapped to a sample mark space;
(307) inputting the result (306) into a softmax classification function to obtain a prediction type probability;
(308) calculating loss by using the output of (307) and the label value, and updating the model parameters;
(309) and performing self-adaptive optimization according to the change trend of the classification accuracy of the verification set in the training process until a convolutional neural network model with classification and recognition capabilities is obtained.
(4) Introducing a target type soft label calculation and training method into the model obtained in the step (3) based on a knowledge distillation technology; the method comprises the following specific steps:
(401) loading the convolutional neural network classification model obtained in the step (3), and adjusting the temperature parameter of the Softmax function;
(402) extracting cross-scene classification sharing spatial features (LOFAR) of data in a target scene for labeling, and using the LOFAR as a training sample of a new neural network;
(403) inputting the time-frequency spectrogram into a convolution layer of a convolution neural network model to extract characteristic information, and enabling the characteristic information to pass through a nonlinear activation function;
(404) sending the output of the nonlinear activation function into a pooling layer, performing feature dimension reduction, and reserving key information;
(405) adjusting corresponding parameters, repeating the steps (403) and (404) for three times, and inputting the result into a full connection layer to enable the characteristics to be mapped to a sample mark space;
(406) inputting the result (405) into a softmax classification function, and extracting a target type soft label;
(407) calculating the loss by using the output of (406) and the label value, and forming a label space and category relation knowledge mapping under a new scene;
(408) and after the training is finished, adjusting the temperature parameter of the Softmax function to be 1, and deploying and applying.
The following is a more specific example:
a cross-scene underwater sound target classification method based on model transfer learning specifically comprises the following steps:
s1: filtering noise reduction
In an actual working environment, no matter source domain data or target domain data, a plurality of targets usually exist in a detection range at the same time, that is, a condition that multi-target features are interleaved exists in data collected by sonar equipment, and deep learning needs feature samples which are as clean as possible, so that a method is needed for suppressing noise signals of other targets as much as possible when signal data of a certain target is obtained. Fig. 2 is a flow chart of filtering and denoising processing in this embodiment, where data needs to be divided at equal intervals, a split beam correlation method is used to perform spectrum peak detection in all directions by using target spatial distribution characteristics to select a target suspected direction, time accumulation is performed on the target suspected direction, an empirical threshold decision is performed to obtain a target accurate direction, time-domain beam forming is performed on the direction to complete spatial filtering, and the purpose of filtering target signals in other directions is achieved; and further performing frequency filtering on the data obtained by the space-domain filtering through an FIR filter, so that the interference of ocean background noise and other noise is reduced, and clean sample data is obtained.
S2: time-frequency feature extraction
Referring to fig. 3, the time-frequency feature extraction process flow is shown, for filtered sonar signal data (L)1(n)，L2(n)，...，Lk(n) performing a short-time Fourier transform to generate a time-frequency spectrum sequence (F)1(u，v)，F2(u，v)，…，Fk(u, v)), the implementation steps are as follows:
s21: by controlling the overlapping of frames, dividing a sampling sequence of an original signal into a plurality of continuous frames which are in front-back connection;
s22: for each frame signal xk(n) performing short-time Fourier transform to obtain Fk(m，n)；
By using a window function, the short-time fourier transform is made to have local characteristics, so that a "local spectrum" of the underwater acoustic signal can be acquired by using the window function, and the operation is as follows:
Figure BDA0003289689540000091
wherein g is*Is a window function, N is a frequency index, N is the number of FFT points, and m is a sliding step length.
And selecting a Hanning window as a window function, obtaining an LOFAR spectrogram through fast Fourier transform, and extracting low-frequency line spectrum characteristics of the acoustic signal.
S3: convolutional neural network model training
A convolutional neural network model of learning time-frequency characteristics is built by referring to the network model structure of FIG. 4, and all sample data are divided into a training set, a verification set and a test set according to the ratio of 3: 1. In the training stage, the feature maps of the training set and the verification set are input into the model, the classification accuracy change of the verification set in the training process is observed, and the hyper-parameters are manually adjusted to improve the generalization capability and robustness of the model. In the testing stage, the characteristic diagram in the test set is input into the trained CNN model, and the training process of the model is completed.
S31: when a CNN model is built based on features of a LOFAR spectrogram, the size of the spectrogram needs to be normalized to 128 × 128, so that the consistency of dimensions is ensured;
s32: inputting the normalized LOFAR spectrogram into a convolution layer with 8 × 8 convolution kernels, 5 channels and 1 step length to obtain an ith hidden layer hi, wherein the working process of the convolution kernels in the convolution layer can be represented by the following formula:
Figure BDA0003289689540000101
(i，j)∈{0，1，...，Ll+1}
Figure BDA0003289689540000102
wherein b is the deviation amount, ZlAnd Zl+1Convolution inputs and outputs representing the l +1 th layer, also called the signature, Wl+1Is a convolution kernel of layer L +1, Ll+1Is Zl+1The feature pattern length and width are assumed to be the same. Z (i, j) corresponds to the pixel of the characteristic diagram, kl is the channel number of the characteristic diagram, k refers to a certain channel, f, s0And p is a convolutional layer parameter, corresponding to the convolutional kernel size, convolutional step size, and number of filling layers.
To increase the non-linear relationship between the layers of the neural network and alleviate the over-fitting problem, hi is input into the ReLU activation function as follows:
Figure BDA0003289689540000103
s33: and (3) sampling and dimensionality reduction is carried out on the feature map by using a pooling layer with a result input window length of 2 and a step length of 2, irrelevant information is removed, the parameter quantity is reduced, and the general representation form of the pooling layer is as follows:
Figure BDA0003289689540000111
step length s in the formula0Pixel (i, j) has the same meaning as the convolution layer, and p is a pre-specified parameter. When p → ∞ the pooling takes a maximum within the region, it is called maximum pooling.
S34: and adjusting the convolution kernel parameters of the convolution layer, continuously iterating for 2 times by adopting a method of the convolution layer + the ReLU function + the pooling layer, and further extracting deep features of the feature map. After all the required local features are obtained, the local features are mapped to a mark space of a sample through 2 full-connection layers in sequence to obtain a one-dimensional vector; using 2 fully connected layers is to better fit the true probability distribution;
s35: and then utilizing a softmax function to represent the multi-classification result in a probability form, wherein the calculation method of the softmax function is as follows:
Figure BDA0003289689540000112
wherein f iscA vector denoted c;
s36: the cross entropy function is used as the loss function, and the concrete form is as follows
Figure BDA0003289689540000113
Wherein t isiIndicates the tag value, yiRepresenting the output of the softmax function. Root of herbaceous plant
S37: an adam (adaptive motion estimation) algorithm is adopted as an optimization algorithm of a deep learning model, and adaptive adjustment of a learning rate is realized by using first-order Moment estimation and second-order Moment estimation, wherein the formula is as follows:
vdw＝0vdb＝0Sdw＝0Sdb＝0
Figure BDA0003289689540000114
Figure BDA0003289689540000115
Figure BDA0003289689540000121
Figure BDA0003289689540000122
the parameter updating formula is as follows:
Figure BDA0003289689540000123
Figure BDA0003289689540000124
the deviation correction is carried out using an exponentially weighted average algorithm in the following way:
Figure BDA0003289689540000125
Figure BDA0003289689540000126
where t denotes the number of iterations, β1Is a parameter of momentum, usually 0.9, beta is a parameter of RMSprop, usually 0.999, e is mainly used to avoid the case where the denominator is 0, usually 10-8W and b represent weights and biases of the neural network, respectively, and α represents a learning rate.
Figure BDA0003289689540000127
The square of the gradient of the parameter w is indicated.
S4: training of target cross-scene underwater sound target classification model
A convolutional neural network of learning time-frequency characteristics is built by referring to a network model structure of FIG. 5, a Softmax function with high temperature is used for reserving enough class relation information, and target domain samples are divided into a training set, a testing set and a verification set according to the proportion of 3: 1. And in the training stage, extracting the target type soft label, and finishing the fitting of the target sample and the label space on the target type relation based on the loss of the relation knowledge between the classes of the soft label. In the testing stage, the characteristic diagram in the test set is input into the trained CNN model, and the training process of the model is completed.
S41: setting the temperature parameter of the Softmax function to be T-5 in the training stage, and calculating the calculation method of the Softmax function according to the multi-classification result presented in the probability form:
Figure BDA0003289689540000131
wherein f iscA vector denoted c, T being a temperature parameter;
s42: extracting soft labels of the category t by calculating a Softmax average value of a target domain type t sample excitation function, and training a new model in the same training process (S3);
s43: after the training is completed, the temperature parameter is adjusted by 1.
According to the method, sonar signal features extracted by a time-frequency analysis method are used as cross-scene classification shared space features, a convolutional neural network is firstly used for deep learning in a source domain to form stable classification capability, knowledge is coded in model parameters, model priori knowledge and a model framework, and then a model-based transfer learning technology is applied to a target scene. In the actual measurement process, a deep neural network model trained by 14600 samples of 4 types of targets collected in offshore sea areas of south China sea is used for verifying the classification accuracy by 90.6%. When the method is transferred to a target scene to process data, the classification accuracy can reach 85% only by a small number of samples (the sampling rate is 5kS/s, the length is 3.768 seconds), the method can be quickly adapted to a new environment and a new task under the conditions of small samples and unbalance, and the time, the energy and the resources consumed by large-amount data acquisition of a target domain and model retraining are greatly saved.
In a word, the knowledge between classes is extracted and migrated by utilizing the strong expression capacity of deep learning, and the knowledge distillation technology is particularly adopted to realize the capacity of cross-domain application of the deep neural network model in the aspect of underwater sound target passive identification, and the classification effect is good. The method is an effective technical approach aiming at the practical problems of high underwater sound data acquisition cost, unbalanced samples and the like.

Claims (6)

1. A cross-scene underwater sound target classification method based on model transfer learning is characterized by comprising the following steps:
(1) filtering and denoising an underwater sound target radiation noise signal received by sonar equipment, filtering out clutter and enhancing the signal to form processed time domain signal data;
(2) performing advanced feature mapping processing on the time domain signal data obtained in the step (1), and extracting a time-frequency spectrogram with inter-class discrimination as cross-scene identification shared space features;
(3) determining hyper-parameters of a convolutional neural network model, marking the class of the time-frequency spectrogram, and training the convolutional neural network model;
(4) introducing a knowledge distillation technology into the convolutional neural network model obtained in the step (3), extracting a target type soft label and training, so that class relation knowledge in the model is fitted to a classification performance target in a new sample and label space;
(5) carrying out cross-scene underwater sound target classification by using the convolutional neural network model trained in the step (4);
and completing the cross-scene underwater sound target classification based on model transfer learning.
2. The cross-scene underwater sound target classification method based on model transfer learning according to claim 1 is characterized in that the specific mode of the step (1) is as follows:
(101) dividing real number form underwater acoustic target radiation noise signals collected by a matrix into equal-time-length samples;
(102) carrying out beam forming on the sample data by adopting a split beam correlation method, and carrying out spectrum peak detection in all directions to select a target suspected direction;
(103) carrying out time accumulation on the suspected target position obtained in the step (102), and obtaining a target accurate position through empirical threshold judgment;
(104) performing time domain beam forming according to the target accurate position obtained in the step (103), and extracting enhanced time domain signal data;
(105) and (5) performing frequency domain filtering on the time domain signal data obtained in the step (104) through an FIR filter to finish filtering and noise reduction.
3. The cross-scene underwater sound target classification method based on model transfer learning according to claim 1, wherein the concrete mode of the step (2) is as follows:
and performing short-time Fourier transform on each frame of time domain signal data, and extracting a time-frequency spectrogram with inter-class discrimination as cross-scene identification shared space characteristics.
4. The model migration learning-based cross-scene underwater sound object classification method according to claim 1, wherein the convolutional neural network model in the step (3) comprises 1 input layer, 4 convolutional layers, 4 maximum pooling layers, 2 full-connected layers and 1 Softmax layer, and the convolutional layers use ReLU function as the activation function.
5. The cross-scene underwater sound target classification method based on model transfer learning of claim 1, wherein in the step (3), the time-frequency spectrogram is labeled to the category to which the time-frequency spectrogram belongs, and a convolutional neural network model is trained in a specific manner of:
(301) labeling the time spectrum graph according to the target type recorded in the collection process, and using the labeled time spectrum graph as a training sample of the convolutional neural network model;
(302) dividing all time-frequency spectrogram characteristic data into a training set, a verification set and a test set, and disordering the sequence to ensure that various types of data in each sample set are uniformly distributed;
(303) and training the convolutional neural network model by using the sample set data to obtain the convolutional neural network model with the classification and recognition capability.
6. The cross-scene underwater sound target classification method based on model transfer learning according to claim 4 is characterized in that the specific method in the step (4) is as follows:
(401) adjusting the temperature parameter of a Softmax function in the convolutional neural network model;
(402) extracting cross-scene classification sharing spatial features of data in a target scene for labeling, and taking the cross-scene classification sharing spatial features as training samples;
(403) inputting the training sample into a convolution layer of a convolution neural network model, and passing a nonlinear activation function through the convolution layer;
(404) sending the output of the nonlinear activation function into a pooling layer, performing feature dimension reduction, and reserving key information;
(405) adjusting corresponding parameters, repeating the steps (403) and (404) for a set number of times, and inputting the result into the full connection layer;
(406) inputting the output result of the full connection layer into a softmax function, and extracting a target type soft tag;
(407) calculating the loss of the target type soft label output in the step (406) and the label value, training the soft label, and forming label space and class relation knowledge mapping under a new scene;
(408) and after the training is finished, adjusting the temperature parameter of the Softmax function to be 1, and deploying and applying.
CN202111160013.1A 2021-09-30 2021-09-30 Cross-scene underwater sound target classification method based on model transfer learning Pending CN113780242A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111160013.1A CN113780242A (en) 2021-09-30 2021-09-30 Cross-scene underwater sound target classification method based on model transfer learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111160013.1A CN113780242A (en) 2021-09-30 2021-09-30 Cross-scene underwater sound target classification method based on model transfer learning

Publications (1)

Publication Number Publication Date
CN113780242A true CN113780242A (en) 2021-12-10

Family

ID=78854661

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111160013.1A Pending CN113780242A (en) 2021-09-30 2021-09-30 Cross-scene underwater sound target classification method based on model transfer learning

Country Status (1)

Country Link
CN (1) CN113780242A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114330441A (en) * 2021-12-28 2022-04-12 厦门大学 Underwater sound JANUS signal identification method and system based on time-frequency spectrum and transfer learning
CN115019149A (en) * 2022-08-03 2022-09-06 中国电子科技集团公司第五十四研究所 Zero sample identification method based on model interpretation result
CN115147718A (en) * 2022-06-21 2022-10-04 北京理工大学 Scene self-adaption system and method for unmanned mobile terminal visual analysis
CN116776230A (en) * 2023-08-22 2023-09-19 北京海格神舟通信科技有限公司 Method and system for identifying signal based on feature imprinting and feature migration
CN116973901A (en) * 2023-09-14 2023-10-31 海底鹰深海科技股份有限公司 Algorithm application of time-frequency analysis in sonar signal processing

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114330441A (en) * 2021-12-28 2022-04-12 厦门大学 Underwater sound JANUS signal identification method and system based on time-frequency spectrum and transfer learning
CN115147718A (en) * 2022-06-21 2022-10-04 北京理工大学 Scene self-adaption system and method for unmanned mobile terminal visual analysis
CN115147718B (en) * 2022-06-21 2024-05-28 北京理工大学 Scene self-adaptive system and method for unmanned mobile terminal visual analysis
CN115019149A (en) * 2022-08-03 2022-09-06 中国电子科技集团公司第五十四研究所 Zero sample identification method based on model interpretation result
CN115019149B (en) * 2022-08-03 2022-11-11 中国电子科技集团公司第五十四研究所 Zero sample identification method based on model interpretation result
CN116776230A (en) * 2023-08-22 2023-09-19 北京海格神舟通信科技有限公司 Method and system for identifying signal based on feature imprinting and feature migration
CN116776230B (en) * 2023-08-22 2023-11-14 北京海格神舟通信科技有限公司 Method and system for identifying signal based on feature imprinting and feature migration
CN116973901A (en) * 2023-09-14 2023-10-31 海底鹰深海科技股份有限公司 Algorithm application of time-frequency analysis in sonar signal processing

Similar Documents

Publication Publication Date Title
CN112364779B (en) Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion
CN113780242A (en) Cross-scene underwater sound target classification method based on model transfer learning
CN109765053B (en) Rolling bearing fault diagnosis method using convolutional neural network and kurtosis index
CN109060001B (en) Multi-working-condition process soft measurement modeling method based on feature transfer learning
CN110245608B (en) Underwater target identification method based on half tensor product neural network
CN109993280B (en) Underwater sound source positioning method based on deep learning
CN109522857B (en) People number estimation method based on generation type confrontation network model
CN111160176B (en) Fusion feature-based ground radar target classification method for one-dimensional convolutional neural network
CN111126386A (en) Sequence field adaptation method based on counterstudy in scene text recognition
CN110197205A (en) A kind of image-recognizing method of multiple features source residual error network
CN105913081B (en) SAR image classification method based on improved PCAnet
CN111126332B (en) Frequency hopping signal classification method based on contour features
CN113392931B (en) Hyperspectral open set classification method based on self-supervision learning and multitask learning
CN107766893B (en) Target identification method based on label multilevel coding neural network
CN114595732B (en) Radar radiation source sorting method based on depth clustering
CN113076920B (en) Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model
CN103077408B (en) Method for converting seabed sonar image into acoustic substrate classification based on wavelet neutral network
CN111275108A (en) Method for performing sample expansion on partial discharge data based on generation countermeasure network
CN108596044B (en) Pedestrian detection method based on deep convolutional neural network
CN111178438A (en) ResNet 101-based weather type identification method
CN112001115B (en) Soft measurement modeling method of semi-supervised dynamic soft measurement network
CN114972904B (en) Zero sample knowledge distillation method and system based on fighting against triplet loss
CN117009916A (en) Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning
CN108694375B (en) Imaging white spirit identification method applicable to multi-electronic nose platform
CN113109782B (en) Classification method directly applied to radar radiation source amplitude sequence

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination