CN111709315A - Underwater acoustic target radiation noise identification method based on field adaptation - Google Patents

Underwater acoustic target radiation noise identification method based on field adaptation Download PDF

Info

Publication number
CN111709315A
CN111709315A CN202010463032.0A CN202010463032A CN111709315A CN 111709315 A CN111709315 A CN 111709315A CN 202010463032 A CN202010463032 A CN 202010463032A CN 111709315 A CN111709315 A CN 111709315A
Authority
CN
China
Prior art keywords
target
domain
label
sample
target domain
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
CN202010463032.0A
Other languages
Chinese (zh)
Inventor
胡桥
于志洋
续丹
郑惠文
刘钰
付同强
唐滢瑾
田芮铭
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.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
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 Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202010463032.0A priority Critical patent/CN111709315A/en
Publication of CN111709315A publication Critical patent/CN111709315A/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
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Abstract

The invention discloses a field-adaptive underwater acoustic target radiation noise identification method, which comprises the steps of constructing a convolutional neural network shared by a source domain and a target domain, pre-training the convolutional neural network by using source domain label samples, respectively calculating source domain classification loss and target domain pseudo label classification loss by using the source domain label samples and the target domain samples, multilayer calculating the multi-core maximum mean difference distance of the source domain label samples and the target domain samples in the convolutional neural network, obtaining network total loss according to the source domain classification loss, the target domain pseudo label classification loss and the multi-core maximum mean difference distance, endowing a network-predicted target domain sample label to a label-free sample through pseudo label learning, enabling the label-free target domain underwater acoustic target data to have the capability of monitoring and training a model, and realizing the mapping of the class characteristics of the source domain underwater acoustic target data and the target domain underwater acoustic target data to the same label space, the problem of accurately identifying the underwater acoustic target sample when the target domain underwater acoustic target sample is not labeled and the data volume is rare is effectively solved.

Description

Underwater acoustic target radiation noise identification method based on field adaptation
Technical Field
The invention belongs to the field of underwater acoustic target detection and identification, and particularly relates to an underwater acoustic target radiation noise identification method based on field adaptation.
Background
In recent years, the exploration of industrial science and technology in the marine field is deepened in various countries in the world, the emphasis on the underwater sound target in the marine field is gradually increased, the research on an accurate and efficient identification method for the underwater sound target becomes more important, and China still lags behind in this respect. Therefore, with the remarkable changes of the international strategic situation and the surrounding safety environment, research on underwater acoustic target identification needs to be advanced.
Document CN201910661350, a big data-based underwater acoustic target intelligent identification method, first collects a large amount of underwater acoustic target original signals, then extracts typical features of the signals manually by means of signal processing and the like, then constructs a training set from the extracted features to train the built neural network model, and finally inputs the training set to obtain an underwater acoustic target classification result. However, the method firstly needs to manually extract features, the automation degree of underwater sound target identification is not high, and end-to-end identification cannot be realized. Meanwhile, when the marginal distribution or conditional distribution of the target recognition domain data and the marginal distribution or conditional distribution used for model training data in the prior art are not completely the same, the method is easy to cause a lower recognition accuracy problem. Due to the interference of complex ocean background noise and the like, the classification method of underwater sound target radiation noise proposed by many scholars in recent years cannot obtain ideal classification performance.
Disclosure of Invention
The invention aims to provide an underwater sound target radiation noise identification method based on field adaptation so as to overcome the defects of the prior art.
In order to achieve the purpose, the invention adopts the following technical scheme:
a method for recognizing radiation noise of an underwater sound target based on field adaptation comprises the following steps:
step 1), constructing a convolutional neural network shared by a source domain and a target domain;
step 2), pre-training the convolutional neural network constructed in the step 1) by using the source domain label sample to obtain a pre-training model, and initializing a target domain sample pseudo label through the pre-training model;
step 3), respectively calculating source domain classification loss and target domain pseudo label classification loss by using the source domain label samples and the target domain samples, simultaneously calculating the cross-domain multi-core maximum mean difference distance of the source domain label samples and the target domain samples in a convolutional neural network in a multi-layer mode, and obtaining network total loss according to the source domain classification loss, the target domain pseudo label classification loss and the multi-core maximum mean difference distance;
and 4) training and optimizing the convolutional neural network aiming at the total network loss obtained in the step 3), continuously updating a target domain sample pseudo label according to the change of the model parameter of the convolutional neural network while training, circularly iterating until the loss function is converged to obtain an optimal underwater acoustic target radiation noise identification model, and identifying the target sample to be detected by adopting the optimal underwater acoustic target radiation noise identification model to obtain an underwater acoustic target radiation noise classification result.
Further, the convolutional neural network comprises 2 convolutional layers, 2 maximum pooling layers, 2 full-link layers and 1 output layer; the 1 st layer and the 3 rd layer of the convolutional neural network are convolutional layers, and the number of convolutional kernels corresponding to each convolutional layer is 10; the 2 nd layer and the 4 th layer of the convolutional neural network are maximum pooling layers, the size of the pooling window of each maximum pooling layer is 1X2, and the step length is 1; the 5 th layer and the 6 th layer of the convolutional neural network are full connection layers, the number of neurons of the 5 th layer is 2480, and the number of neurons of the 6 th layer is 256; layer 7 is the output layer and the activation function is the softmax function.
Further, in the step 2), the source domain label sample is input into the convolutional neural network constructed in the step 1), iterative optimization is carried out on the convolutional neural network model parameters through an optimization algorithm until the loss function is converged, and a pre-training model is obtained.
Further, in step 2), inputting the unlabeled target domain sample into a convolutional neural network, then outputting the probability that the unlabeled target domain sample is in each category, and taking the category corresponding to the maximum probability as the target domain sample pseudo label of the unlabeled target domain sample; when the target domain sample pseudo label is initialized, all non-label samples in the target domain sample are input into a pre-training model, each non-label sample is subjected to forward propagation to obtain a network output layer neuron value, and the output layer neuron value is converted into probability distribution of corresponding sample types through a softmax function.
Further, the method specifically comprises the following steps: the softmax function is expressed as follows:
Figure BDA0002511684500000031
wherein e is a natural index, FiIs the i-th neuron value of the output layer, SiTaking the probability value of the i-th class label for the label-free sample;
then obtaining the probability distribution of the target domain samples in the mark space: p ═ S1S2…Sk]；
Wherein k is the total number of categories;
and finally, outputting the target domain sample pseudo label through the following formula:
Figure BDA0002511684500000032
Figure BDA0002511684500000033
wherein the argmax function returns the index of the maximum value in the array,
Figure BDA0002511684500000034
is a binarized pseudo label for the target domain sample.
Further, a cross entropy function is adopted as a loss function of the source domain label samples, and the source domain classification loss is as follows:
Figure BDA0002511684500000035
wherein the content of the first and second substances,
Figure BDA0002511684500000036
the probability of the prediction of the ith sample of the source domain being a true label,
Figure BDA0002511684500000037
the true label of the corresponding source domain label exemplar, nSThe number of samples of the source domain small batch of the input network in each iteration period.
Further, a cross entropy function is adopted as a loss function in the target domain sample training, and the target domain pseudo label classification loss is as follows:
Figure BDA0002511684500000041
wherein the content of the first and second substances,
Figure BDA0002511684500000042
the probability of being a true label for the prediction of the jth sample of the target domain,
Figure BDA0002511684500000043
pseudo-label of the target domain sample for the jth sample of the target domain, nTThe number of samples of the target domain small batch input into the network in each iteration period is counted.
Further, the difference distance of the maximum mean values of the multiple nuclei is as follows:
Figure BDA0002511684500000044
wherein, XS、XTRespectively a source domain and a target domainIs characterized in that n and m are the small-batch sample numbers of the source domain and the target domain respectively,
Figure BDA00025116845000000411
representation-to-core mapping
Figure BDA0002511684500000045
The relevant feature mapping, H is the regenerated kernel Hilbert space,
Figure BDA00025116845000000412
is defined as a convex combination of a plurality of kernel functions, wherein:
Figure BDA0002511684500000046
wherein, βuAs a weighting coefficient, kuIs a Gaussian kernel function, and m is the number of kernel functions;
the total loss of the network is:
Figure BDA0002511684500000047
wherein the content of the first and second substances,
Figure BDA0002511684500000048
in order to classify the loss for the source domain,
Figure BDA0002511684500000049
for the target domain pseudo label classification loss, α, β are the hyper-parameters set before training, L is the set composed of network convolution layer and full link layer,
Figure BDA00025116845000000410
output characteristics, k, of the source domain and the target domain of the l-th layer of the network, respectivelylFor the compromise factor:
Figure BDA0002511684500000051
further, the training process of the convolutional neural network comprises the following steps:
1. calculating network total loss using chain-type derivative rule
Figure BDA0002511684500000052
The model parameters are adjusted by the initial learning rate of 0.001 and the learning rate is continuously decreased;
2. updating the target domain sample pseudo label according to the adjusted model parameter;
3. and (3) continuously repeating the step 1 to the step 2, and circularly iterating until the loss function is converged to finally obtain the optimal underwater sound target radiation noise identification model.
Compared with the prior art, the invention has the following beneficial technical effects:
the invention discloses a field-adaptive underwater acoustic target radiation noise identification method, which comprises the steps of constructing a convolutional neural network shared by a source domain and a target domain, pre-training the convolutional neural network by using source domain label samples, respectively calculating source domain classification loss and target domain pseudo label classification loss by using the source domain label samples and the target domain samples, simultaneously calculating multi-core maximum mean difference distance of the source domain label samples and the target domain samples in a multi-layer mode in the convolutional neural network, obtaining network total loss according to the source domain classification loss, the target domain pseudo label classification loss and the multi-core maximum mean difference distance, initializing target domain sample pseudo labels through a pre-training model, endowing the non-labeled samples with network-predicted target domain sample labels through pseudo label learning, and enabling the non-labeled target domain underwater acoustic target data to have the capability of a supervision training model, finally, mapping of the classification features of the source domain underwater acoustic target data and the target domain underwater acoustic target data to the same label space is achieved, simultaneous adaptation of cross-domain marginal probability distribution and conditional probability distribution of the source domain data and the target domain data is achieved through multi-core maximum mean difference and pseudo label learning, the problem of accurate identification of the target domain underwater acoustic target samples when the target domain underwater acoustic target samples are not labeled and the data volume is small is effectively solved, and finally the classification performance of the target identification field underwater acoustic target is improved.
Furthermore, from the perspective of field adaptation, a convolutional neural network is constructed, MK-MMD is used as a measurement criterion of marginal distribution difference of a source domain and a target domain, MK-MMD distance is adapted to multiple layers in a depth network at the same time, the test performance of the convolutional neural network is superior to that of an MMD field adaptation method widely used at present, the marginal distribution distance between cross domains is effectively reduced, and accumulated source domain underwater acoustic target label data is applied to underwater acoustic target data identification of a target attention field to realize target identification.
Further, the source domain classification loss and the target domain pseudo label classification loss are respectively calculated through forward propagation and are merged into the network total loss for training so as to draw the condition distribution difference between the source domain and the target domain.
Furthermore, original bandwidth parameters in the kernel function are replaced by calculating the square of the average value difference of the characteristics of the input source domain and the target domain, the problems that the experimental performance fluctuation is large and the like caused by different artificially set Gaussian kernel function bandwidth parameters in the calculation of the MK-MMD are solved, and the MK-MMD distance is fused into the total loss of the network to draw the boundary distribution difference across the domains.
Drawings
FIG. 1 is a flow chart in an embodiment of the present invention.
Fig. 2 is a specific structural diagram of a convolutional neural network in an embodiment of the present invention.
Fig. 3 is a diagram of the total loss of the network according to the method of the present invention.
Fig. 4 is a schematic diagram of five types of underwater acoustic targets related to a training data set in the embodiment of the present invention.
FIG. 5 is a visualization diagram of sample migration characteristics t-SNE of a source domain and a target domain after domain adaptation in the embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the accompanying drawings:
the following detailed description of specific embodiments of the present invention is made with reference to the accompanying drawings and examples. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention.
As shown in fig. 1, the method for identifying radiation noise of underwater acoustic target based on field adaptation of the present invention includes the following steps:
step 1: constructing a convolutional neural network shared by a source domain and a target domain;
in the method of the present invention, the specific structure of the constructed convolutional neural network is shown in fig. 2. The convolutional neural network comprises 7 layers, wherein the 7 layers comprise 2 convolutional layers, 2 maximum pooling layers, 2 full-connection layers and 1 output layer; the layer 1 and the layer 3 of the convolutional neural network are convolutional layers, the number of convolutional kernels corresponding to the convolutional layers is 10, the sizes of the convolutional kernels are 1X5 and 1X3 respectively, the step length is 1, and the activation function is a Leaky _ ReLU function;
the 2 nd layer and the 4 th layer of the convolutional neural network are maximum pooling layers, the size of the pooling window of each maximum pooling layer is 1X2, and the step length is 1; the 5 th and 6 th layers of the convolutional neural network are full connection layers, the number of neurons of the 5 th layer is 2480, the number of neurons of the 6 th layer is 256, and the activation function is a Leaky _ ReLU function; layer 7 is the output layer and the activation function is the softmax function.
The method and the device share the same implementation field of parameters of all network layers (convolutional layers and full-connection layers) when the source domain label sample and the target domain sample are transmitted in the forward direction of the network. The field adaptation is realized by constructing a convolution neural network shared by a source field and a target field, namely, the aim of migrating the knowledge of the source field data to the target field is fulfilled by drawing the distribution difference of the source field data and the target field data in the same characteristic space. In short, the related knowledge is obtained in a field learning, and the application to a new field is completed through a field adaptation method. Therefore, different from the traditional deep learning, the training samples of the field adaptation method are divided into source field label data and target field label-free data, the source field label data is used for solving the problems that the sample size is rare and no reliable label exists in the target identification field, and the method is the final target of the method.
2, pre-training the convolutional neural network constructed in the step 1 by using a source domain label sample to obtain a pre-training model, and initializing a target domain sample pseudo label through the pre-training model;
the training set of the convolutional neural network is composed of source domain label samples and target domain samples without labels.
In each experiment, the training set comprises source domain label samples and target domain samples, wherein the used source domain label samples and target domain samples are respectively from radiated noise signals collected by the same sonar collecting equipment at corresponding ports when the 5-class underwater acoustic targets are in different sailing states. The schematic diagram of 5-class underwater sound target types is shown in fig. 4, and the schematic diagram is respectively a motorboat, a pilot ship, a passenger ship, an ocean-going passenger ship and a ro-ro ship. 1095 source domain label samples are collected, wherein each ship type comprises 219 samples, and each target domain sample comprises 485 samples, wherein each ship type comprises 97 samples; each sample is a radiated noise signal of length 1000 points, and is a one-dimensional time series.
And (3) during pre-training, inputting the source domain label sample into the convolutional neural network constructed in the step (1), and performing iterative optimization on the model parameters of the convolutional neural network through an optimization algorithm until the loss function is converged to obtain a pre-training model.
Inputting an unlabeled target domain sample into a convolutional neural network, then outputting the probability that the unlabeled target domain sample is in each category, and taking the category corresponding to the maximum probability as a target domain sample pseudolabel of the unlabeled target domain sample; when the target domain sample pseudo label is initialized, all non-label samples in the target domain sample are input into a pre-training model, each non-label sample is subjected to forward propagation to obtain a network output layer neuron value, and the output layer neuron value is converted into probability distribution of corresponding sample types through a softmax function. The softmax function is expressed as follows:
Figure BDA0002511684500000081
wherein e is a natural index, FiIs the i-th neuron value of the output layer, SiTaking the probability value of the i-th class label for the label-free sample;
then obtaining the probability distribution of the target domain samples in the mark space: p ═ S1S2…Sk]；
Wherein k is the total number of categories;
and finally, outputting the target domain sample pseudo label through the following formula to finish initialization of the target domain sample pseudo label:
Figure BDA0002511684500000091
Figure BDA0002511684500000092
wherein the argmax function returns the index of the maximum value in the array,
Figure BDA0002511684500000093
is a binarized pseudo label for the target domain sample.
And step 3: respectively calculating source domain classification loss and target domain pseudo label classification loss by using a source domain label sample and a target domain sample, simultaneously calculating the cross-domain multi-core maximum mean difference distance of the source domain label sample and the target domain sample in a convolutional neural network in a multi-layer mode, and obtaining network total loss according to the source domain classification loss, the target domain pseudo label classification loss and the multi-core maximum mean difference distance;
in the example, a cross entropy function is adopted as a loss function when a source domain label sample and a target domain sample are trained; the source domain classification loss calculation is as follows:
Figure BDA0002511684500000094
wherein the content of the first and second substances,
Figure BDA0002511684500000095
the probability of the prediction of the ith sample of the source domain being a true label,
Figure BDA0002511684500000096
the true label of the corresponding source domain label exemplar, nSInputting the source domain small batch sample number of the network in each iteration period;
the target domain pseudo label classification loss calculation formula is as follows:
Figure BDA0002511684500000097
wherein the content of the first and second substances,
Figure BDA0002511684500000098
the probability of being a true label for the prediction of the jth sample of the target domain,
Figure BDA0002511684500000099
pseudo-label of the target domain sample for the jth sample of the target domain, nTThe number of samples of the target domain small batch input into the network in each iteration period is counted.
Adapting the cross-domain probability distribution to obtain the total network loss, and inputting the source domain label sample into a pre-training model to obtain the source domain classification loss; and inputting the target domain sample into a pre-training model according to the initialized target domain sample pseudo label to obtain the target domain pseudo label classification loss.
The cross-domain probability distribution of the source domain data and the target domain data comprises marginal distribution and conditional distribution;
the adaptation of the condition distribution means that under the condition that sample class labels are considered, the migration characteristics of the source domain label samples and the target domain samples are mapped to the same label space by simultaneously optimizing the source domain classification loss and the target domain pseudo label classification loss. When the conditional distribution of cross-domain data is specifically adapted, the source domain label samples and the target domain samples are simultaneously input into a pre-training model, the source domain classification loss and the target domain pseudo label classification loss are respectively calculated through forward propagation, and the total network loss for training is integrated to draw the conditional distribution difference between the source domain and the target domain.
The marginal distribution difference of the cross-domain data refers to the distribution distance of the features of all samples in the source domain data and the target domain data in a high-dimensional space without considering the sample class labels. The adaptation to the boundary distribution is achieved by minimizing the MK-MMD between the source domain label samples and the target domain samples in the training. In domain adaptation, since the marginal distributions of the source domain data and the target domain data are different, in order to enable adaptation between the cross-domain data, the distribution difference between them needs to be evaluated: the multi-core maximum mean difference (MK-MMD) is characterized in that different domain samples are evenly embedded into a regeneration core Hilbert space (RKHS) by fusing a plurality of appropriate kernel functions, and then the marginal distribution difference between cross domains is measured. Meanwhile, in order to solve the problems of large experimental performance fluctuation and the like caused by different artificially set Gaussian kernel function bandwidth parameters in MK-MMD calculation, a median heuristic bandwidth estimation method is selected, and original bandwidth parameters in the kernel function are replaced by calculating the square of the average value difference of input source domain and target domain characteristics. Calculating the MK-MMD distance of cross-domain in a convolutional neural network in a multi-layer mode, and integrating the MK-MMD distance into the total loss of the network to draw up the marginal distribution difference of the cross-domain;
the expression for the MK-MMD distance is as follows:
Figure BDA0002511684500000111
wherein, Xs、XTRespectively the characteristics of a source domain and a target domain, n and m respectively represent the small batch sample numbers of the source domain and the target domain,
Figure BDA0002511684500000119
representation-to-core mapping
Figure BDA0002511684500000112
The relevant feature mapping, H is the regenerated kernel Hilbert space,
Figure BDA00025116845000001110
defined as a convex combination of a plurality of kernel functions, expressed as follows:
Figure BDA0002511684500000113
wherein, βuAs a weighting coefficient, kuIs a gaussian kernel function, and m is the number of kernel functions.
Finally, as shown in fig. 3, the total loss of the network is:
Figure BDA0002511684500000114
wherein the content of the first and second substances,
Figure BDA0002511684500000115
in order to classify the loss for the source domain,
Figure BDA0002511684500000116
for the target domain pseudo label classification loss, α and β are hyper-parameters set before training, L is a set formed by the 1 st, 3 rd, 5 th and 6 th layers of the network,
Figure BDA0002511684500000117
output characteristics, k, of the source domain and the target domain of the l-th layer of the network, respectivelylFor the compromise factor, it is as follows:
Figure BDA0002511684500000118
and 4, step 4: and (3) training and optimizing the network total loss convolutional neural network obtained in the step (3), continuously updating the pseudo label of the target domain sample according to the change of the model parameter of the convolutional neural network while training, circularly iterating until the loss function is converged to obtain an optimal underwater acoustic target radiation noise identification model, and identifying the target sample to be detected by adopting the optimal underwater acoustic target radiation noise identification model to obtain an underwater acoustic target radiation noise classification result.
In this example, an Adam optimizer is used that updates the model parameters in the convolutional neural network by calculating an update step by comprehensively considering the first moment estimate and the second moment estimate of the gradient. The method has the advantages of simple implementation, high calculation efficiency, less memory requirement, no influence of gradient expansion transformation on updating of model parameters, good interpretive performance of hyper-parameters, no adjustment or little fine adjustment in general and the like.
The training process of the convolutional neural network comprises the following steps:
1) calculating the total loss of the network by using the chain-type derivation rule
Figure BDA0002511684500000121
A gradient of (a); adjusting the model parameters by continuously decreasing the initial learning rate of 0.001;
2) updating the target domain sample pseudo label according to the adjusted model parameter;
3) and continuously repeating the step 1) to the step 2), and circularly iterating until the loss function is converged to finally obtain the optimal underwater sound target radiation noise identification model.
Specifically, the target domain samples are input into the optimal underwater acoustic target radiation noise identification model obtained through training in the step 4, and then the underwater acoustic target radiation noise classification result can be obtained.
And (3) inputting all target domain samples serving as a test set into the network trained in the step 5, outputting 5 values by each sample through forward propagation, wherein each value represents the probability that the sample to be tested belongs to each class, and the higher the value is, the higher the probability that the sample to be tested belongs to the corresponding underwater sound target class is, so that the classification purpose is achieved. For example, an output vector of (0.1, 0.2, 0.1, 0.0, 0.6) represents that the probability that the target domain sample to be measured belongs to the category 5 underwater sound target class is the largest.
Finally, as shown in fig. 5, migration features of source domain and target domain samples extracted at a layer before an output layer are visualized through t-SNE, and it is proved that the method provided by the invention can effectively adapt to distribution differences among different domain data to improve target domain data prediction performance.
In summary, the invention provides a field adaptation-based underwater acoustic target radiation noise identification method aiming at the problem of different probability distributions of source domain data and target domain data in the traditional underwater acoustic target identification field, and compared with the traditional underwater acoustic target identification technology, the method provided by the invention better adapts marginal distribution and conditional distribution between the source domain data and the target domain data, effectively solves the problem of accurately identifying target domain underwater acoustic target samples when the target domain underwater acoustic target samples are not labeled and the data amount is rare, and finally improves the classification performance of the target identification field underwater acoustic targets. The classification method is adopted to perform classification experiments on five different types of underwater acoustic target radiation noise signals, and compared with two methods based on parameter migration and MMD field adaptation, experimental results prove that the accuracy of underwater acoustic target identification performed by the method reaches 93.8%, the highest accuracy of the other two migration methods using the same data set is 81.7%, and the accuracy is remarkably improved, so that the method can better perform classification identification on the underwater acoustic target radiation noise signals of ships and warships and the like.
The above-mentioned contents are only for explaining the technical idea of the invention of the present application, and can not be used as the basis for limiting the protection scope of the invention, and any modifications and substitutions made on the technical solution according to the design concept and technical features proposed by the present invention are within the protection scope of the claims of the present invention.

Claims (9)

1. A method for identifying underwater sound target radiation noise based on field adaptation is characterized by comprising the following steps:
step 1), constructing a convolutional neural network shared by a source domain and a target domain;
step 2), pre-training the convolutional neural network constructed in the step 1) by using the source domain label sample to obtain a pre-training model, and initializing a target domain sample pseudo label through the pre-training model;
step 3), respectively calculating source domain classification loss and target domain pseudo label classification loss by using the source domain label samples and the target domain samples, simultaneously calculating the cross-domain multi-core maximum mean difference distance of the source domain label samples and the target domain samples in a convolutional neural network in a multi-layer mode, and obtaining network total loss according to the source domain classification loss, the target domain pseudo label classification loss and the multi-core maximum mean difference distance;
and 4) training and optimizing the convolutional neural network aiming at the total network loss obtained in the step 3), continuously updating a target domain sample pseudo label according to the change of the model parameter of the convolutional neural network while training, circularly iterating until the loss function is converged to obtain an optimal underwater acoustic target radiation noise identification model, and identifying the target sample to be detected by adopting the optimal underwater acoustic target radiation noise identification model to obtain an underwater acoustic target radiation noise classification result.
2. The method for recognizing the radiation noise of the underwater acoustic target based on the field adaptation as claimed in claim 1, wherein the convolutional neural network comprises 2 convolutional layers, 2 max pooling layers, 2 full-link layers and 1 output layer; the 1 st layer and the 3 rd layer of the convolutional neural network are convolutional layers, and the number of convolutional kernels corresponding to each convolutional layer is 10; the 2 nd layer and the 4 th layer of the convolutional neural network are maximum pooling layers, the size of the pooling window of each maximum pooling layer is 1X2, and the step length is 1; the 5 th layer and the 6 th layer of the convolutional neural network are full connection layers, the number of neurons of the 5 th layer is 2480, and the number of neurons of the 6 th layer is 256; layer 7 is the output layer and the activation function is the softmax function.
3. The method for recognizing the radiation noise of the underwater acoustic target based on the field adaptation as claimed in claim 1, wherein in the step 2), the source domain label sample is input into the convolutional neural network constructed in the step 1), and iterative optimization is performed on the convolutional neural network model parameters through an optimization algorithm until the loss function is converged to obtain a pre-training model.
4. The method for identifying the underwater acoustic target radiation noise based on the field adaptation as claimed in claim 1, wherein in the step 2), the unlabeled target domain sample is input into a convolutional neural network, then the probability that the unlabeled target domain sample is in each category is output, and the category corresponding to the maximum probability value is taken as the target domain sample pseudo label of the unlabeled target domain sample; when the target domain sample pseudo label is initialized, all non-label samples in the target domain sample are input into a pre-training model, each non-label sample is subjected to forward propagation to obtain a network output layer neuron value, and the output layer neuron value is converted into probability distribution of corresponding sample types through a softmax function.
5. The method for identifying the underwater acoustic target radiation noise based on the field adaptation as claimed in claim 4, is characterized in that: the softmax function is expressed as follows:
Figure FDA0002511684490000021
wherein e is a natural index, FiIs the i-th neuron value of the output layer, SiTaking the probability value of the i-th class label for the label-free sample;
then obtaining the probability distribution of the target domain samples in the mark space: p ═ S1S2…Sk]；
Wherein k is the total number of categories;
and finally, outputting the target domain sample pseudo label through the following formula:
Figure FDA0002511684490000022
Figure FDA0002511684490000023
wherein the argmax function returns the index of the maximum value in the array,
Figure FDA0002511684490000024
is a binarized pseudo label for the target domain sample.
6. The method for recognizing the radiation noise of the underwater acoustic target based on the field adaptation as claimed in claim 1, wherein a cross entropy function is adopted as a loss function of the source domain label samples, and the source domain classification loss is as follows:
Figure FDA0002511684490000031
wherein the content of the first and second substances,
Figure FDA0002511684490000032
the probability of the prediction of the ith sample of the source domain being a true label,
Figure FDA0002511684490000033
the true label of the corresponding source domain label exemplar, nSThe number of samples of the source domain small batch of the input network in each iteration period.
7. The method for identifying the underwater acoustic target radiation noise based on the field adaptation as claimed in claim 1, wherein a cross entropy function is adopted as a loss function in the target domain sample training, and the target domain pseudo label classification loss is as follows:
Figure FDA0002511684490000034
wherein the content of the first and second substances,
Figure FDA0002511684490000035
the probability of being a true label for the prediction of the jth sample of the target domain,
Figure FDA0002511684490000036
pseudo-label of the target domain sample for the jth sample of the target domain, nTThe number of samples of the target domain small batch input into the network in each iteration period is counted.
8. The method for recognizing the radiation noise of the underwater acoustic target based on the field adaptation as claimed in claim 1, wherein the difference distance of the maximum mean of the multiple cores is as follows:
Figure FDA0002511684490000037
wherein, XS、XTRespectively the characteristics of a source domain and a target domain, n and m respectively represent the small batch sample numbers of the source domain and the target domain,
Figure FDA0002511684490000038
representation-to-core mapping
Figure FDA0002511684490000039
Figure FDA00025116844900000310
The relevant feature mapping, H is the regenerated kernel Hilbert space,
Figure FDA00025116844900000311
is defined as a convex combination of a plurality of kernel functions, wherein:
Figure FDA0002511684490000041
wherein, βuAs a weighting coefficient, kuIs a Gaussian kernel function, and m is the number of kernel functions;
the total loss of the network is:
Figure FDA0002511684490000042
wherein the content of the first and second substances,
Figure FDA0002511684490000043
in order to classify the loss for the source domain,
Figure FDA0002511684490000044
for the target domain pseudo label classification loss, α, β are the hyper-parameters set before training, L is the set composed of network convolution layer and full link layer,
Figure FDA0002511684490000045
output characteristics, k, of the source domain and the target domain of the l-th layer of the network, respectivelylFor the compromise factor:
Figure FDA0002511684490000046
9. the method for recognizing the radiation noise of the underwater acoustic target based on the field adaptation as claimed in claim 1, wherein the training process of the convolutional neural network comprises the following steps:
1. calculating network total loss using chain-type derivative rule
Figure FDA0002511684490000047
The model parameters are adjusted by the initial learning rate of 0.001 and the learning rate is continuously decreased;
2. updating the target domain sample pseudo label according to the adjusted model parameter;
3. and (3) continuously repeating the step 1 to the step 2, and circularly iterating until the loss function is converged to finally obtain the optimal underwater sound target radiation noise identification model.
CN202010463032.0A 2020-05-27 2020-05-27 Underwater acoustic target radiation noise identification method based on field adaptation Pending CN111709315A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010463032.0A CN111709315A (en) 2020-05-27 2020-05-27 Underwater acoustic target radiation noise identification method based on field adaptation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010463032.0A CN111709315A (en) 2020-05-27 2020-05-27 Underwater acoustic target radiation noise identification method based on field adaptation

Publications (1)

Publication Number Publication Date
CN111709315A true CN111709315A (en) 2020-09-25

Family

ID=72538604

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010463032.0A Pending CN111709315A (en) 2020-05-27 2020-05-27 Underwater acoustic target radiation noise identification method based on field adaptation

Country Status (1)

Country Link
CN (1) CN111709315A (en)

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112270208A (en) * 2020-09-28 2021-01-26 天津大学 Cross data domain micro-representation classification method based on feature subspace separation
CN112507621A (en) * 2020-12-16 2021-03-16 电子科技大学 Radar interference semi-supervised identification method based on label propagation
CN112580518A (en) * 2020-12-22 2021-03-30 杭州电子科技大学 Cross-tested EEG cognitive state identification method based on prototype clustering domain adaptive algorithm
CN112669331A (en) * 2020-12-25 2021-04-16 上海交通大学 Target data migration iterative learning method and target data migration iterative learning system
CN112686333A (en) * 2021-01-19 2021-04-20 科润智能控制股份有限公司 Switch cabinet partial discharge mode identification method based on depth subdomain adaptive migration network
CN112861616A (en) * 2020-12-31 2021-05-28 电子科技大学 Passive field self-adaptive target detection method
CN113095381A (en) * 2021-03-29 2021-07-09 西安交通大学 Underwater sound target identification method and system based on improved DBN
CN113168567A (en) * 2018-10-29 2021-07-23 Hrl实验室有限责任公司 System and method for small sample transfer learning
CN113269024A (en) * 2021-03-31 2021-08-17 中南大学 Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics
CN113360659A (en) * 2021-07-19 2021-09-07 云南大学 Cross-domain emotion classification method and system based on semi-supervised learning
CN113420775A (en) * 2021-03-31 2021-09-21 中国矿业大学 Image classification method under extremely small quantity of training samples based on adaptive subdomain field adaptation of non-linearity
CN113449631A (en) * 2021-06-25 2021-09-28 中南大学 Image classification method and system
CN113902104A (en) * 2021-11-01 2022-01-07 南京工程学院 Non-invasive load monitoring method combining unsupervised domain self-adaptive strategy and attention mechanism
CN113989595A (en) * 2021-11-05 2022-01-28 西安交通大学 Federal multi-source domain adaptation method and system based on shadow model
CN114444605A (en) * 2022-01-30 2022-05-06 南京邮电大学 Unsupervised domain adaptation method based on double-unbalance scene
WO2022193973A1 (en) * 2021-03-17 2022-09-22 腾讯科技（深圳）有限公司 Image processing method and apparatus, electronic device, computer readable storage medium, and computer program product
CN115471662A (en) * 2022-11-03 2022-12-13 深圳比特微电子科技有限公司 Training method, recognition method, device and storage medium of semantic segmentation model
CN116108353A (en) * 2023-04-12 2023-05-12 厦门大学 Small sample deep learning underwater sound target recognition method based on data packet
CN117113061A (en) * 2023-09-14 2023-11-24 中国人民解放军军事科学院系统工程研究院 Cross-receiver radiation source fingerprint identification method and device

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109948741A (en) * 2019-03-04 2019-06-28 北京邮电大学 A kind of transfer learning method and device
CN110321926A (en) * 2019-05-24 2019-10-11 北京理工大学 A kind of moving method and system based on depth residual GM network

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109948741A (en) * 2019-03-04 2019-06-28 北京邮电大学 A kind of transfer learning method and device
CN110321926A (en) * 2019-05-24 2019-10-11 北京理工大学 A kind of moving method and system based on depth residual GM network

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
BIN YANG等: ""A Transfer Learning Method for Intelligent Fault Diagnosis from Laboratory Machines to Real-Case Machines"", 《2018 INTERNATIONAL CONFERENCE ON SENSING, DIAGNOSTICS, PROGNOSTICS, AND CONTROL》 *
SHAN LI等: ""Deep Emotion Transfer Network for Cross-database Facial Expression Recognition"", 《2018 24TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR)》 *
周牧等: ""基于多核最大均值差异迁移学习的WLAN室内入侵检测方法"", 《电子与信息学报》 *
桂存斌: ""基于最小化最大平均差异损失的无监督领域自适应"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *
王翎等: ""结合域混淆与MK-MMD的深度适应网络"", 《小型微型计算机系统》 *
秦磊: ""基于深度学习的通信辐射源个体识别"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *

Cited By (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113168567A (en) * 2018-10-29 2021-07-23 Hrl实验室有限责任公司 System and method for small sample transfer learning
CN112270208B (en) * 2020-09-28 2023-07-07 天津大学 Cross data domain micro-expression classification method based on feature subspace separation
CN112270208A (en) * 2020-09-28 2021-01-26 天津大学 Cross data domain micro-representation classification method based on feature subspace separation
CN112507621A (en) * 2020-12-16 2021-03-16 电子科技大学 Radar interference semi-supervised identification method based on label propagation
CN112507621B (en) * 2020-12-16 2023-04-07 电子科技大学 Radar interference semi-supervised identification method based on label propagation
CN112580518B (en) * 2020-12-22 2024-04-05 杭州电子科技大学 Cross-test EEG cognitive state recognition method based on prototype clustering domain adaptation algorithm
CN112580518A (en) * 2020-12-22 2021-03-30 杭州电子科技大学 Cross-tested EEG cognitive state identification method based on prototype clustering domain adaptive algorithm
CN112669331A (en) * 2020-12-25 2021-04-16 上海交通大学 Target data migration iterative learning method and target data migration iterative learning system
CN112861616A (en) * 2020-12-31 2021-05-28 电子科技大学 Passive field self-adaptive target detection method
CN112686333A (en) * 2021-01-19 2021-04-20 科润智能控制股份有限公司 Switch cabinet partial discharge mode identification method based on depth subdomain adaptive migration network
WO2022193973A1 (en) * 2021-03-17 2022-09-22 腾讯科技（深圳）有限公司 Image processing method and apparatus, electronic device, computer readable storage medium, and computer program product
CN113095381A (en) * 2021-03-29 2021-07-09 西安交通大学 Underwater sound target identification method and system based on improved DBN
CN113095381B (en) * 2021-03-29 2024-04-05 西安交通大学 Underwater sound target identification method and system based on improved DBN
CN113420775A (en) * 2021-03-31 2021-09-21 中国矿业大学 Image classification method under extremely small quantity of training samples based on adaptive subdomain field adaptation of non-linearity
CN113269024B (en) * 2021-03-31 2022-04-26 中南大学 Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics
CN113269024A (en) * 2021-03-31 2021-08-17 中南大学 Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics
CN113420775B (en) * 2021-03-31 2024-03-29 中国矿业大学 Picture classification method based on nonlinearity self-adaptive subdomain field adaptation under extremely small quantity of training samples
CN113449631A (en) * 2021-06-25 2021-09-28 中南大学 Image classification method and system
CN113360659A (en) * 2021-07-19 2021-09-07 云南大学 Cross-domain emotion classification method and system based on semi-supervised learning
CN113902104A (en) * 2021-11-01 2022-01-07 南京工程学院 Non-invasive load monitoring method combining unsupervised domain self-adaptive strategy and attention mechanism
CN113989595A (en) * 2021-11-05 2022-01-28 西安交通大学 Federal multi-source domain adaptation method and system based on shadow model
CN114444605B (en) * 2022-01-30 2024-04-05 南京邮电大学 Unsupervised domain adaptation method based on double unbalanced scene
CN114444605A (en) * 2022-01-30 2022-05-06 南京邮电大学 Unsupervised domain adaptation method based on double-unbalance scene
CN115471662B (en) * 2022-11-03 2023-05-02 深圳比特微电子科技有限公司 Training method, recognition method, device and storage medium for semantic segmentation model
CN115471662A (en) * 2022-11-03 2022-12-13 深圳比特微电子科技有限公司 Training method, recognition method, device and storage medium of semantic segmentation model
CN116108353A (en) * 2023-04-12 2023-05-12 厦门大学 Small sample deep learning underwater sound target recognition method based on data packet
CN117113061B (en) * 2023-09-14 2024-02-23 中国人民解放军军事科学院系统工程研究院 Cross-receiver radiation source fingerprint identification method and device
CN117113061A (en) * 2023-09-14 2023-11-24 中国人民解放军军事科学院系统工程研究院 Cross-receiver radiation source fingerprint identification method and device

Similar Documents

Publication Publication Date Title
CN111709315A (en) Underwater acoustic target radiation noise identification method based on field adaptation
CN109993280B (en) Underwater sound source positioning method based on deep learning
CN109086824B (en) Seabed substrate sonar image classification method based on convolutional neural network
CN110245608B (en) Underwater target identification method based on half tensor product neural network
CN112364779B (en) Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion
CN109828251B (en) Radar target identification method based on characteristic pyramid light-weight convolution neural network
CN110109059B (en) Radar radiation source signal identification method based on deep learning network
CN112001270B (en) Ground radar automatic target classification and identification method based on one-dimensional convolutional neural network
CN110728360A (en) Micro-energy device energy identification method based on BP neural network
Yang et al. Underwater acoustic research trends with machine learning: general background
CN111160176A (en) Fusion feature-based ground radar target classification method for one-dimensional convolutional neural network
CN110991257B (en) Polarized SAR oil spill detection method based on feature fusion and SVM
CN109581339B (en) Sonar identification method based on automatic adjustment self-coding network of brainstorming storm
CN114595732B (en) Radar radiation source sorting method based on depth clustering
CN110991516A (en) Side-scan sonar image target classification method based on style migration
CN113780242A (en) Cross-scene underwater sound target classification method based on model transfer learning
Ye et al. Recognition algorithm of emitter signals based on PCA+ CNN
CN113064133B (en) Sea surface small target feature detection method based on time-frequency domain depth network
CN113076920B (en) Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model
Wang et al. Ship radiated noise recognition technology based on ML-DS decision fusion
CN113221758A (en) Underwater acoustic target identification method based on GRU-NIN model
CN113111786A (en) Underwater target identification method based on small sample training image convolutional network
CN115114949A (en) Intelligent ship target identification method and system based on underwater acoustic signals
CN114814776B (en) PD radar target detection method based on graph attention network and transfer learning
CN113095381B (en) Underwater sound target identification method and system based on improved DBN

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20200925