CN117911852A - Underwater target distance prediction method based on self-adaption in part of unsupervised field - Google Patents
Underwater target distance prediction method based on self-adaption in part of unsupervised field Download PDFInfo
- Publication number
- CN117911852A CN117911852A CN202410319967.XA CN202410319967A CN117911852A CN 117911852 A CN117911852 A CN 117911852A CN 202410319967 A CN202410319967 A CN 202410319967A CN 117911852 A CN117911852 A CN 117911852A
- Authority
- CN
- China
- Prior art keywords
- training
- model
- data set
- sound pressure
- simulation
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 81
- 238000012549 training Methods 0.000 claims abstract description 226
- 238000004088 simulation Methods 0.000 claims abstract description 116
- 230000006870 function Effects 0.000 claims abstract description 78
- 238000005259 measurement Methods 0.000 claims abstract description 60
- 230000003044 adaptive effect Effects 0.000 claims abstract description 37
- 238000013136 deep learning model Methods 0.000 claims abstract description 31
- 239000011159 matrix material Substances 0.000 claims description 26
- 238000010606 normalization Methods 0.000 claims description 21
- 238000005070 sampling Methods 0.000 claims description 19
- 238000012545 processing Methods 0.000 claims description 18
- 230000006978 adaptation Effects 0.000 claims description 12
- 238000000605 extraction Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000013508 migration Methods 0.000 abstract description 7
- 230000005012 migration Effects 0.000 abstract description 7
- 230000000694 effects Effects 0.000 abstract description 6
- 230000008569 process Effects 0.000 description 10
- 238000002474 experimental method Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 5
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 4
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 4
- 238000001228 spectrum Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000008485 antagonism Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000002542 deteriorative effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/05—Underwater scenes
-
- 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/30—Assessment of water resources
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
Abstract
The invention discloses an adaptive underwater target distance prediction method based on a part of unsupervised field, which comprises the following steps: establishing an actual measurement training data set and a simulation training data set; pre-training the deep learning model by using a simulation training data set to obtain a data pre-training model; constructing a domain self-adaptive model by combining the data pre-training model and the domain self-adaptive network, and performing distance prediction and domain prediction training on the domain self-adaptive model by utilizing the actual measurement training data set and the simulation training data set; constructing a first loss function of distance prediction training and a second loss function of domain prediction training, and weighting to obtain a third loss function; and updating the domain self-adaptive model by using a third loss function to obtain an underwater target distance prediction model for distance prediction. According to the invention, through weighting the loss function, the actually measured data are aligned with the simulation data with the same distance range, so that the negative migration effect generated by overlarge distance between the data and the distance range is relieved, and the underwater target distance prediction performance is improved.
Description
Technical Field
The invention relates to the technical field of ship and ocean engineering, in particular to an adaptive underwater target distance prediction method based on a part of unsupervised field.
Background
In recent years, the technology of detecting underwater targets has been gaining more and more attention in the fields of ocean resource development, ocean science research and the like. The traditional underwater target detection technology generally adopts a matching field processing method to detect the target, but the traditional matching field processing method has poor universal prediction performance, so the prior art provides an underwater sound source distance prediction method capable of extracting statistical characteristics from actual received signals to construct a parameter estimation model on the basis of a deep learning technology.
In order to solve the problem, the prior art proposes a method for predicting by using simulation data and actual measurement data, which does not depend on data labeling, but needs to ensure that the distance range covered by the simulation data and the actual measurement data is approximately consistent, if the distance between the two covered distance ranges is too large, the positioning performance is greatly reduced due to the negative migration effect, which greatly restricts the application of the method in the actual scene. Therefore, how to alleviate the negative migration effect generated by the too large data distance range gap and improve the prediction performance of the underwater target distance prediction is a problem to be solved by those skilled in the art.
Disclosure of Invention
The embodiment of the invention provides an adaptive underwater target distance prediction method based on a part of unsupervised field, aiming at improving the prediction performance of the underwater target distance prediction.
The embodiment of the invention provides an underwater target distance prediction method based on self-adaption in a part of unsupervised field, which comprises the following steps:
Acquiring an actual measurement complex sound pressure data set and a simulation complex sound pressure data set of an underwater target, setting a label for the simulation complex sound pressure data set, and then respectively establishing an actual measurement training data set and a simulation training data set based on the actual measurement complex sound pressure data set and the simulation complex sound pressure data set with the label;
pre-training a pre-built deep learning model by using the simulation training data set to construct a data pre-training model;
Constructing a domain self-adaptive model by combining the data pre-training model and a domain self-adaptive network, performing distance prediction training on the domain self-adaptive model by utilizing the actual measurement training data set and the simulation training data set, and performing domain prediction training on the domain self-adaptive model by utilizing the actual measurement training data set and the simulation training data set;
respectively constructing a first loss function related to distance prediction training and a second loss function related to domain prediction training, and weighting the first loss function and the second loss function to obtain a third loss function;
Carrying out parameter updating on the domain self-adaptive model by utilizing the third loss function so as to construct and obtain an underwater target distance prediction model;
and acquiring complex sound pressure data of the appointed underwater target, and performing distance prediction on the acquired complex sound pressure data through the underwater target distance prediction model.
The embodiment of the invention firstly uses the simulation training data set to pretrain a deep learning model, builds a data pretraining model, secondly combines the data pretraining model and a domain self-adaptive network to build a domain self-adaptive model, uses the actual measurement training data set and the simulation training data set to conduct distance prediction training and domain prediction training on the domain self-adaptive model, builds a first loss function related to the distance prediction training and a second loss function related to the domain prediction training, then weights the first loss function and the second loss function to obtain a third loss function, then uses the third loss function to update parameters of the domain self-adaptive model, builds an underwater target distance prediction model, and finally conducts distance prediction through the underwater target distance prediction model. According to the embodiment of the invention, the loss function is weighted, so that the actually measured data and the simulation data with the same distance range are distributed and aligned, and the negative migration effect generated by overlarge data distance range difference is relieved, and the prediction performance of the underwater target distance prediction is improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic flow chart of an adaptive underwater target distance prediction method based on a part of unsupervised field according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of model training based on a partially unsupervised domain-based adaptive underwater target distance prediction method according to an embodiment of the present invention;
FIG. 3 is a schematic sub-flowchart of an adaptive underwater target distance prediction method based on a part of the unsupervised field according to an embodiment of the present invention;
fig. 4 is a first experimental example diagram of an underwater target distance prediction method based on partial unsupervised field adaptation according to an embodiment of the present invention;
FIG. 5 is a diagram of a second experimental example of a method for predicting a distance between underwater targets based on partial unsupervised domain adaptation according to an embodiment of the present invention;
fig. 6 is a third experimental example diagram of an underwater target distance prediction method based on partial unsupervised field adaptation according to an embodiment of the present invention;
fig. 7 is a diagram of a fourth experimental example of an underwater target distance prediction method based on partial unsupervised field adaptation according to an embodiment of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be understood that the terms "comprises" and "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.
It is also to be understood that the terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
It should be further understood that the term "and/or" as used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes such combinations.
Referring to fig. 1 and fig. 2, fig. 1 is a schematic flow chart of an underwater target distance prediction method based on partial unsupervised field adaptation according to an embodiment of the present invention, which specifically includes: steps S101-S106.
S101, acquiring an actual measurement complex sound pressure data set and a simulation complex sound pressure data set of an underwater target, setting a label for the simulation complex sound pressure data set, and then respectively establishing an actual measurement training data set and a simulation training data set based on the actual measurement complex sound pressure data set and the simulation complex sound pressure data set with the label;
s102, pre-training a pre-built deep learning model by using the simulation training data set to construct a data pre-training model;
S103, constructing a domain self-adaptive model by combining the data pre-training model and a domain self-adaptive network, performing distance prediction training on the domain self-adaptive model by utilizing the actual measurement training data set and the simulation training data set, and performing domain prediction training on the domain self-adaptive model by utilizing the actual measurement training data set and the simulation training data set;
s104, respectively constructing a first loss function related to distance prediction training and a second loss function related to domain prediction training, and weighting the first loss function and the second loss function to obtain a third loss function;
s105, carrying out parameter updating on the domain self-adaptive model by utilizing the third loss function so as to construct and obtain an underwater target distance prediction model;
s106, acquiring complex sound pressure data of the appointed underwater target, and performing distance prediction on the acquired complex sound pressure data through the underwater target distance prediction model.
In this embodiment, firstly, an actual measurement training data set and a simulation training data set are respectively built according to an actual measurement complex sound pressure data set and a simulation complex sound pressure data set of an underwater target, secondly, a deep learning model is built and pre-trained by utilizing the simulation training data set to build a data pre-training model, then a domain self-adaptive model is built by combining the data pre-training model and a domain self-adaptive network, the actual measurement training data set and the simulation training data set are respectively input into the domain self-adaptive model to perform distance prediction training and domain prediction training, so that a first loss function corresponding to the distance prediction training and a second loss function corresponding to the domain prediction training are respectively built, the first loss function and the second loss function are weighted, the domain self-adaptive model is subjected to parameter updating by utilizing a third loss function obtained by weighting, the underwater target distance prediction model is built, and finally, the underwater target specified by the underwater target distance prediction model is subjected to distance prediction.
According to the embodiment, the loss function is weighted, so that the measured data can be aligned with the simulation data distribution with the same distance range, and the measured data is not aligned with the simulation data distribution with different distance ranges, thereby relieving the negative migration effect caused by overlarge data distance range gap, and improving the prediction performance of the underwater target distance prediction.
In one embodiment, the step S101 includes:
Obtaining a real distance corresponding to each simulated complex sound pressure sample in the simulated complex sound pressure data set, and distributing labels to each simulated complex sound pressure sample according to the following steps of:
.
wherein, The distance segment length is represented, y represents a label, r represents a real distance corresponding to the simulated complex sound pressure sample, u represents the distance segment number, r max represents a maximum value in the real distances corresponding to all the simulated complex sound pressure samples, and r min represents a minimum value in the real distances corresponding to all the simulated complex sound pressure samples.
In this embodiment, a real distance corresponding to each simulated complex sound pressure sample in the simulated complex sound pressure data set is first obtained, and then label distribution is performed on each simulated complex sound pressure sample based on the corresponding real distance, so as to construct a simulated complex sound pressure data set with labels. It will be appreciated that the simulated training data set subsequently created based on the simulated complex sound pressure data set with the tag also has the tag. In a specific application scene, the length of the distance segment can be set according to actual requirementsAnd a distance segment number u, for example, a distance segment length/>, may be set100M, and a distance segment number u of 93.
Referring to fig. 3, in one embodiment, the step S101 includes steps S301 to S304.
S301, respectively performing short-time Fourier transform on all complex sound pressure samples in the actually measured complex sound pressure data set and the simulated complex sound pressure data set to obtain broadband complex sound pressures respectively corresponding to all the complex sound pressure samples;
s302, carrying out normalization processing on each broadband complex sound pressure according to the following formula:
.
wherein, Representing the broadband complex sound pressure,/>The result of normalization processing of the broadband complex sound pressure is shown, L represents the number of the array hydrophones used for collecting complex sound pressure samples, and p j (f) represents the broadband complex sound pressure of the complex sound pressure samples collected by the jth array hydrophone;
s303, constructing a signal sampling covariance matrix for each complex sound pressure sample according to the following formula:
.
wherein C (f) represents the sampling covariance matrix, z represents the snapshot number, H represents the complex conjugate operation, Representing the normalization processing result of the broadband complex sound pressure corresponding to the ith snapshot complex sound pressure sample;
s304, setting the signal sampling covariance matrix of each complex sound pressure sample as a complex sound pressure sample characteristic extraction result, and respectively establishing and obtaining the actual measurement training data set and the simulation training data set based on the complex sound pressure sample characteristic extraction result.
In this embodiment, feature extraction needs to be performed on all the complex sound pressure samples in the actually measured complex sound pressure data set and the simulated complex sound pressure data set, so as to establish an actually measured training data set and a simulated training data set.
Taking the establishment of an actual measurement training data set as an example, firstly performing short-time Fourier transform on all actual measurement complex sound pressure samples in the actual measurement complex sound pressure data set to obtain corresponding frequency spectrum characteristics through time domain signal transformation of the complex sound pressure samples, taking the frequency spectrum characteristics of each complex sound pressure sample as corresponding broadband complex sound pressure, secondly performing normalization processing on each broadband complex sound pressure so as to weaken the influence of a sound source, then constructing a signal sampling covariance matrix based on the normalization result, finally setting the signal sampling covariance matrix of each complex sound pressure sample as a corresponding complex sound pressure sample characteristic extraction result, and establishing the actual measurement training data set corresponding to the actual measurement complex sound pressure data set based on the extraction result.
Similarly, for the simulation training data set, firstly, performing short-time fourier transform on all simulation complex sound pressure samples in the simulation complex sound pressure data set to obtain corresponding frequency spectrum characteristics through time domain signal transformation of the complex sound pressure samples, taking the frequency spectrum characteristics of each complex sound pressure sample as corresponding broadband complex sound pressure, secondly, performing normalization processing on each broadband complex sound pressure so as to weaken the influence of a sound source, then constructing a signal sampling covariance matrix based on the normalization result, finally setting the signal sampling covariance matrix of each complex sound pressure sample as a corresponding complex sound pressure sample characteristic extraction result, and establishing the simulation training data set corresponding to the simulation complex sound pressure data set based on the extraction result.
In a specific application scene, the actual measurement complex sound pressure sample is acquired through the array hydrophone, and after the actual measurement complex sound pressure sample is acquired, short-time Fourier transform and normalization processing are needed according to the calculation method so as to obtain the actual measurement broadband complex sound pressure. The simulated complex sound pressure sample can be generated by the sound field propagation software Kraken, so that the simulated complex sound pressure sample can be subjected to short-time fourier transform and normalization processing according to the calculation method to obtain simulated broadband complex sound pressure, and the sound field propagation software Kraken can also directly generate the broadband complex sound pressure after normalization processing. Referring to fig. 4, for example, the simulated complex sound pressure sample may be generated by using sound field propagation software Kraken according to the terrestrial sound parameter of S5 voyage (fifth voyage in the experiment) in a swelex-96 experiment (a hydroacoustic experiment, which is mainly used to process the target distance estimation problem of the sea area with only a small amount of hydroacoustic data), and the normalization processing result of the broadband complex sound pressure corresponding to the complex sound pressure sample may be directly generated in sound field propagation software Kraken. In addition, one snapshot parameter in the normalization process may be set according to the actual application scene, for example, 2 seconds of data of the complex sound pressure sample may be selected as one snapshot in the normalization process, and the overlapping rate between snapshots may be set to be 50%.
In one embodiment, the step S304 includes:
Normalizing the signal sampling covariance matrix according to the following formula:
.
wherein, An ith channel representing the covariance matrix of the signal samples,/>Minimum value channel representing covariance matrix of signal samples,/>Representing the maximum value channel of the signal sampling covariance matrix.
In this embodiment, since the signal sampling covariance matrix is a complex matrix, normalization processing is required for the signal sampling covariance matrix in order to enable the matrix to be used as input data for subsequent model training. In a specific application scene, the real part and the imaginary part of the signal sampling covariance matrix of F frequency points are respectively placed in a first dimension, and normalization processing is carried out on the signal sampling covariance matrix in the first dimension, so that the data obtained by final normalization processing is a matrix with 2F multiplied by L in dimension, wherein L represents the number of array hydrophones.
In one embodiment, the step S102 includes:
Inputting simulation training samples in the simulation training data set into a generator of the deep learning model, and outputting first simulation data features through a convolution layer of the generator;
and inputting the first simulation data characteristics into a classifier of the deep learning model, classifying by a multi-layer perceptron of the classifier, and outputting classification results as prediction results corresponding to simulation training samples to construct the data pre-training model.
In this embodiment, a deep learning model is built by combining a generator and a classifier, simulation training samples in a simulation training data set are input into the generator of the deep learning model, a convolution layer in the generator is utilized to carry out convolution processing on the simulation training samples, first simulation data features are output, then the first simulation data features are input into the classifier of the deep learning model, the first simulation data features are classified through a multi-layer perceptron of the classifier to obtain a classification result, and finally the classification result is output as a prediction result corresponding to the simulation training samples, so that a data pre-training model is built.
In a specific application scenario, a deep learning model can be designed and built based on the residual module, and the deep learning model is designed to have a generator and a classifier. The deep learning model generator is provided with three convolution layers, wherein the first convolution layer of the generator is a3 multiplied by 3 group convolution layer, the output channel number of the group convolution layer is 66, two convolution layers with the output channel number of 128 are sequentially connected after the group convolution layer, and the output of the group convolution layer is subjected to linear mapping through the second convolution layer and then is added with the output of the third convolution layer by corresponding elements, so that the first simulation data characteristic is obtained. The classifier of the deep learning model consists of two multi-layer perceptrons, all of which adopt random clipping layers as regularization means, and the clipping probability is set to be 50%. Further, the learning rate of the deep learning model pre-training may be set to 0.001.
In an embodiment, the step S102 further includes:
obtaining labels corresponding to each simulation training sample of the simulation training data set, and training the deep learning model by taking the labels as supervision signals;
calculating a training loss function of the deep learning model according to the following formula:
.
Wherein L C represents the training loss function, L CE represents the cross entropy loss function, D s represents the simulated training dataset, c i represents the i-th simulated training sample, y i represents the label corresponding to simulated training sample c i, Representing the prediction result corresponding to the simulation training sample,/>Representing a desire to find;
And updating the deep learning model by using the training loss function, and constructing to obtain the data pre-training model.
In this embodiment, first, the labels corresponding to all the simulation training samples in the simulation training data set are obtained, the labels are used as supervisory signals of the deep learning model to train, and a training loss function of the deep learning model is calculated, so that parameter update and/or weight update are performed on the deep learning model, and a data pre-training model is constructed. In a specific application scenario, the parameters and/or weights of the deep learning model may be updated using a back propagation algorithm.
In one embodiment, the step S103 includes:
generating corresponding first measured data features and second simulation data features for the measured training data set and the simulation training data set respectively through a generator in the data pre-training model;
Performing general feature learning on the first measured data feature and the second simulation data feature by using a domain adaptive network;
And carrying out distance prediction on the first measured data characteristic and the second simulation data characteristic which are subjected to general characteristic learning through a classifier in the data pre-training model to obtain a distance prediction result, so as to construct a first sub-domain self-adaptive model.
In another embodiment, the step S103 further includes:
Generating corresponding second measured data features and third simulation data features for the measured training data set and the simulation training data set respectively through a generator in the data pre-training model;
performing general feature learning on the second measured data feature and the third simulation data feature by using a domain adaptive network;
inputting the second measured data characteristic and the third simulation data characteristic after the general characteristic learning to a gradient inversion layer for countermeasure learning;
Inputting the second measured data characteristic and the third simulation data characteristic after the countermeasure learning to a domain discriminator for domain prediction training, and outputting a domain prediction result to construct a second sub-domain self-adaptive model;
And combining the first sub-domain self-adaptive model and the second sub-domain self-adaptive model to construct the domain self-adaptive model.
In this embodiment, the domain adaptive network is added to the data pre-training model, so that training is performed by combining the domain adaptive network and the data pre-training model, and the first sub-domain adaptive model is constructed. Specifically, firstly, an actual measurement training data set and a simulation training data set are input into a data pre-training model, a generator in the data pre-training model is utilized to generate corresponding first actual measurement data features and second simulation data features, then the first actual measurement data features and the second simulation data features are input into a domain self-adaptive network in the data pre-training model to perform general feature learning, and finally the first actual measurement data features and the second simulation data features after general feature learning are input into a classifier in the data pre-training model to perform distance prediction so as to obtain distance prediction results corresponding to the actual measurement training data set and the simulation training data set, so that a first sub-domain self-adaptive model is constructed.
The embodiment further adds the domain self-adaptive network, the gradient inversion layer and the domain discriminant into the data pre-training model, so that the domain self-adaptive network, the gradient inversion layer, the domain discriminant and the data pre-training model are combined for training, and a second sub-domain self-adaptive model is constructed. Specifically, firstly, an actual measurement training data set and a simulation training data set are input into a data pre-training model, a generator in the data pre-training model is utilized to generate corresponding second actual measurement data features and third simulation data features, secondly, the second actual measurement data features and the third simulation data features are input into a domain adaptive network in the data pre-training model to perform general feature learning, then the second actual measurement data features and the third simulation data features after feature learning are input into a gradient inversion layer of the pre-training model to perform countermeasure learning, and then the second actual measurement data features and the third simulation data features after countermeasure learning are input into a domain discriminator in the data pre-training model to perform domain prediction training, and domain prediction results corresponding to the obtained actual measurement training data set and the simulation training data set are output, so that a second sub-domain adaptive model is constructed.
After the first sub-domain self-adaptive model and the second sub-domain self-adaptive model are respectively constructed, the first sub-domain self-adaptive model and the second sub-domain self-adaptive model are combined to construct the domain self-adaptive model.
According to the embodiment, the domain adaptive network, the gradient inversion layer and the domain arbiter are combined to train the data pre-training model, and the antagonism learning is completed through the gradient inversion layer, so that the statistical distribution distance of the simulation data and the actually measured data is reduced, the domain adaptive network is updated towards the direction of increasing the loss of the domain arbiter, the domain arbiter cannot distinguish the simulation data and the actually measured data, and once the domain arbiter fails, the simulation data and the actually measured data cannot be distinguished, the simulation data and the actually measured data can be considered to be mapped to similar characteristics, and distribution alignment is completed. Specifically, when updating the domain adaptive network, the gradient from the domain arbiter is multiplied by a negative number, so as to perform gradient inversion, thereby completing the countermeasure learning.
In one embodiment, the domain adaptive network is formed by a 128-channel convolutional layer, and the domain arbiter is formed by two multi-layer perceptrons. In addition, the nonlinear mapping functions in all models adopt modified linear units as activation functions, and the corresponding regularization methods adopt batch normalization layers, and the optimizers are set as random gradient descent optimizers. In a specific application scenario, the learning rate of the domain adaptive network may be set to 0.001, and the learning rate of the domain arbiter may be set to 0.01.
In one embodiment, the step S104 includes:
acquiring a first loss function of the first sub-domain adaptive model;
Calculating a second loss function of the second sub-domain adaptive model according to the following formula:
.
Wherein L D denotes the second loss function, Representing a calculation expectation, wherein P s (c) represents the joint distribution of simulation training samples, P t (x) represents the joint distribution of actual measurement training samples, w represents a weight corresponding to the actual measurement training samples, D represents the generator, G represents the discriminator, c i represents the ith simulation training sample, and x i represents the ith actual measurement training sample;
Weighting the first loss function and the second loss function according to the following formula, thereby obtaining the third loss function:
.
Wherein L represents the third loss function, L C represents the first loss function, A trade-off factor representing the first and second loss functions.
In this embodiment, a first loss function corresponding to the first sub-domain adaptive model is calculated first, and a second loss function of the second sub-domain adaptive model is calculated, and then a third loss function for parameter updating of the domain adaptive model is obtained by weighting the first loss function and the second loss function, so that semantic characteristics of data features in the model updating process are ensured. In a specific application scenario, the calculation method of the first loss function is the same as the calculation method of the training loss function pre-trained by the deep learning model.
In one embodiment, the weights corresponding to the measured training samples are calculated according to the following equation:
.
Wherein w (x i) represents a weight corresponding to the ith actual measurement training sample, n t represents the number of the simulation training samples, n represents the number of the actual measurement training samples, And representing the prediction result corresponding to the actual measurement training sample.
In one embodiment, the step S105 includes:
And updating parameters of the domain adaptive model according to the following steps:
.
wherein, Model parameters after updating parameters of the domain adaptive model are represented,/>Representing model parameters before parameter updating of the domain adaptive model,/>Representing learning rate,/>Representing the gradient trade-off factor.
In a specific embodiment, the gradient trade-off factor is calculated as follows:
.
where p represents the number of training rounds for parameter update. The gradient trade-off factor can smooth the training of the domain arbiter, and in the parameter updating process, the gradient trade-off factor can increase along with the increase of the training round number from a value close to 0 until the gradient value is close to 1, so that the gradient value in the early training stage can show a smaller value, the updating amplitude of the countermeasure training is smaller, and in the later training stage, when the training of the model is relatively stable, the larger gradient value can only enable the countermeasure training to be effective.
In another specific embodiment, after the underwater target distance prediction model is constructed, the underwater target distance prediction model is tested by using the actually measured training data set, the test standard adopts the absolute average error MAPE and the reliable prediction probability PCL to calculate, and a specific calculation formula can be:
.
wherein, As the prediction result of the underwater target distance prediction model,/>Is true value (actual measurement training data set), N is the number of samples,/>The value is 1 or 0, which is used for indicating whether the positioning error is smaller than 10% or larger than 10%.
In a specific application scenario, an experiment may be performed based on the underwater target distance prediction method provided in this embodiment, and the experimental result is compared with the experimental result in the prior art to verify the prediction performance of the underwater target distance prediction method provided in this embodiment, and specifically may be compared with a matching field method (MFP, matched-field processing) using a butler processor and an unsupervised field adaptive method (UDA, unsupervised Domain Adaptation) that does not use weighting.
The data partitioning method of the experiment is shown in table 1, wherein three sets of data located in different distance segments are selected to test the effectiveness of the underwater target distance prediction method provided in this embodiment on any distance range. In each scene, the selected measured data covers a distance range of 1000m, while the simulated data covers a distance range of 7100m, the measured data coverage is about 14%. The selected domain adaptation samples in the three scenes do not overlap each other and come from different distance ranges. In the comparison method, the UDA method is identical to training parameters such as the learning rate, the training round number, the migration learning weight and the like of the method, and the MFP method adopts simulation data in table 1 as a copying field. The basic machine learning method of the underwater target distance prediction method provided in this embodiment is based on a partial unsupervised domain adaptive method (PDA, partial domain adaptation), so that the PDA is used as an abbreviation of the underwater target distance prediction method provided in this embodiment for convenience of presentation.
The positioning results of the experiments of table 1 are shown in fig. 5, wherein (a) - (c) are scene 1, (d) - (f) are scene 2, and (g) - (i) are scene 3. As can be seen from the positioning results, the PDA method provided by the embodiment has better neighborhood consistency and fewer predicted outliers, and the positioning effect is better than that of the UDA method and the MFP method under all three conditions.
TABLE 1
。
Table 2 shows the index of positioning performance quantization in table 1. As can be seen from table 2, the PDA method provided in this embodiment achieves the best performance on the MAPE index in all three scenes. It should be noted that the PDA method provided in this embodiment has 0% PCL index in scene 2 and 3, and 0% PCL index in scene 2 in MFP method, while the UDA method has the best performance in both scenes, however, the actual prediction performance of the UDA method is more outlier and the prediction consistency is worse, and the slightly higher PCL index is caused by random prediction, which can be seen from the worse MAPE index of the UDA method compared with the other two methods. Therefore, the algorithm performance cannot be evaluated here by the PCL index alone. For the MFP method and the PDA method provided in this embodiment, the occurrence of 0% of PCL index may be caused by that the predicted distance and the actual distance do not completely coincide due to mismatch of environmental parameters or errors in GPS positioning, but the predicted results all show a distribution conforming to the actual ship running track, so that the errors do not affect the evaluation of the method performance.
TABLE 2
。
Fig. 6 shows the training stability of UDA versus the PDA method provided in this example, where (a) and (c) are the results of the PDA provided in this example, and (b) and (d) are the results of the UDA method. It is also necessary to investigate whether the PDA method provided in this embodiment can keep performance from deteriorating during training, because in practice it is very difficult to determine on which training round training is stopped. Here, all experiments in table 1 were still tested, in each of which 30% of the measured data was selected as a validation set to monitor the training process and UDA method was used as a control. As can be seen from the comparison results, the PDA method provided in this embodiment achieves the stability of training, in each scene, even if a large number of rounds of training are performed after the performance is optimal, the performance remains stable, no dip phenomenon occurs, the UDA method has a gradual degradation phenomenon of MAPE or PCL index in each scene, and the optimal performance of all rounds cannot reach the level of the PDA method provided in this embodiment, which also proves the effectiveness of the PDA method provided in this embodiment.
Fig. 7 is a weight visualization result of the PDA method according to the present embodiment, and fig. 7 is a graph showing a predicted distance vector of the classifier for all measured data (i.e.) Arranged together, the role of weighting domain adaptation can be seen from fig. 7. In fig. 7, the probability distribution of the output category of the PDA method provided in this embodiment shows higher confidence in the interval corresponding to the true distance, while the confidence in the other distance intervals is relatively lower. Because the PDA method provided in this embodiment uses the mean value of the confidence coefficient as the domain adaptive weight, the weight to which the relevant sample is allocated is far greater than that of the irrelevant sample, which allows the PDA method provided in this embodiment to accurately reduce the statistical distribution of the relevant sample and the actually measured data, and to avoid the influence of the irrelevant sample to a certain extent, so that the negative migration phenomenon is avoided in the domain adaptive process, and the final distance prediction performance is improved.
In the description, each embodiment is described in a progressive manner, and each embodiment is mainly described by the differences from other embodiments, so that the same similar parts among the embodiments are mutually referred. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section. It should be noted that it will be apparent to those skilled in the art that various modifications and adaptations of the application can be made without departing from the principles of the application and these modifications and adaptations are intended to be within the scope of the application as defined in the following claims.
It should also be noted that in this specification, relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
Claims (10)
1. The method for predicting the distance of the underwater target based on the self-adaption of the part of the unsupervised field is characterized by comprising the following steps:
Acquiring an actual measurement complex sound pressure data set and a simulation complex sound pressure data set of an underwater target, setting a label for the simulation complex sound pressure data set, and then respectively establishing an actual measurement training data set and a simulation training data set based on the actual measurement complex sound pressure data set and the simulation complex sound pressure data set with the label;
pre-training a pre-built deep learning model by using the simulation training data set to construct a data pre-training model;
Constructing a domain self-adaptive model by combining the data pre-training model and a domain self-adaptive network, performing distance prediction training on the domain self-adaptive model by utilizing the actual measurement training data set and the simulation training data set, and performing domain prediction training on the domain self-adaptive model by utilizing the actual measurement training data set and the simulation training data set;
respectively constructing a first loss function related to distance prediction training and a second loss function related to domain prediction training, and weighting the first loss function and the second loss function to obtain a third loss function;
Carrying out parameter updating on the domain self-adaptive model by utilizing the third loss function so as to construct and obtain an underwater target distance prediction model;
and acquiring complex sound pressure data of the appointed underwater target, and performing distance prediction on the acquired complex sound pressure data through the underwater target distance prediction model.
2. The method for predicting the distance of an underwater target based on the self-adaption of the part of the unsupervised field according to claim 1, wherein the step of setting a tag on the simulated complex sound pressure data set comprises the steps of:
Obtaining a real distance corresponding to each simulated complex sound pressure sample in the simulated complex sound pressure data set, and distributing labels to each simulated complex sound pressure sample according to the following steps of:
.
wherein, The distance segment length is represented, y represents a label, r represents a real distance corresponding to the simulated complex sound pressure sample, u represents the distance segment number, r max represents a maximum value in the real distances corresponding to all the simulated complex sound pressure samples, and r min represents a minimum value in the real distances corresponding to all the simulated complex sound pressure samples.
3. The method for predicting the distance between underwater targets based on the partial unsupervised field adaptation according to claim 1, wherein the establishing the actual measurement training data set and the simulation training data set based on the actual measurement complex sound pressure data set and the simulation complex sound pressure data set with the tag respectively comprises:
respectively carrying out short-time Fourier transform on all complex sound pressure samples in the actually measured complex sound pressure data set and the simulated complex sound pressure data set to obtain broadband complex sound pressures respectively corresponding to all the complex sound pressure samples;
And carrying out normalization processing on each broadband complex sound pressure according to the following steps:
.
wherein, Representing the broadband complex sound pressure,/>The result of normalization processing of the broadband complex sound pressure is shown, L represents the number of the array hydrophones used for collecting complex sound pressure samples, and p j (f) represents the broadband complex sound pressure of the complex sound pressure samples collected by the jth array hydrophone;
Constructing a signal sampling covariance matrix for each complex sound pressure sample according to the normalization processing result:
.
wherein C (f) represents the sampling covariance matrix, z represents the snapshot number, H represents the complex conjugate operation, Representing the normalization processing result of the broadband complex sound pressure corresponding to the ith snapshot complex sound pressure sample;
setting the signal sampling covariance matrix of each complex sound pressure sample as a complex sound pressure sample characteristic extraction result, and respectively establishing and obtaining the actual measurement training data set and the simulation training data set based on the complex sound pressure sample characteristic extraction result.
4. The method for predicting the distance between underwater targets based on the self-adaption of the partially unsupervised field according to claim 3, wherein the setting the signal sampling covariance matrix of each complex sound pressure sample as a complex sound pressure sample feature extraction result and respectively establishing and obtaining the actually measured training data set and the simulated training data set based on the complex sound pressure sample feature extraction result comprises:
Normalizing the signal sampling covariance matrix according to the following formula:
.
wherein, An ith channel representing the covariance matrix of the signal samples,/>Minimum value channel representing covariance matrix of signal samples,/>Representing the maximum value channel of the signal sampling covariance matrix.
5. The method for predicting the distance between underwater targets based on the self-adaption of the partially unsupervised field according to claim 1, wherein the pre-training the pre-built deep learning model by using the simulation training data set to construct a data pre-training model comprises:
Inputting simulation training samples in the simulation training data set into a generator of the deep learning model, and outputting first simulation data features through a convolution layer of the generator;
and inputting the first simulation data characteristics into a classifier of the deep learning model, classifying by a multi-layer perceptron of the classifier, and outputting classification results as prediction results corresponding to simulation training samples to construct the data pre-training model.
6. The method for predicting the distance between underwater targets based on the self-adaption of the partially unsupervised field according to claim 5, wherein the pre-training the pre-built deep learning model by using the simulation training data set to construct a data pre-training model, further comprises:
obtaining labels corresponding to each simulation training sample of the simulation training data set, and training the deep learning model by taking the labels as supervision signals;
calculating a training loss function of the deep learning model according to the following formula:
.
Wherein L C represents the training loss function, L CE represents the cross entropy loss function, D s represents the simulated training dataset, c i represents the i-th simulated training sample, y i represents the label corresponding to simulated training sample c i, Representing the prediction result corresponding to the simulation training sample,/>Representing a desire to find;
And updating the deep learning model by using the training loss function, and constructing to obtain the data pre-training model.
7. The method for predicting the distance of an underwater target based on the self-adaption of the partially unsupervised field according to claim 1, wherein the constructing a domain self-adaption model by combining the data pre-training model and a domain self-adaption network, performing distance prediction training on the domain self-adaption model by using the actual measurement training data set and the simulation training data set, and performing domain prediction training on the domain self-adaption model by using the actual measurement training data set and the simulation training data set comprises:
generating corresponding first measured data features and second simulation data features for the measured training data set and the simulation training data set respectively through a generator in the data pre-training model;
Performing general feature learning on the first measured data feature and the second simulation data feature by using a domain adaptive network;
And carrying out distance prediction on the first measured data characteristic and the second simulation data characteristic which are subjected to general characteristic learning through a classifier in the data pre-training model to obtain a distance prediction result, so as to construct a first sub-domain self-adaptive model.
8. The method for predicting the distance of an underwater target based on the self-adaption in the partially unsupervised field according to claim 7, wherein the constructing a domain self-adaption model by combining the data pre-training model and a domain self-adaption network and performing distance prediction training on the domain self-adaption model by using the actual measurement training data set and the simulation training data set and performing domain prediction training on the domain self-adaption model by using the actual measurement training data set and the simulation training data set further comprises:
Generating corresponding second measured data features and third simulation data features for the measured training data set and the simulation training data set respectively through a generator in the data pre-training model;
performing general feature learning on the second measured data feature and the third simulation data feature by using a domain adaptive network;
inputting the second measured data characteristic and the third simulation data characteristic after the general characteristic learning to a gradient inversion layer for countermeasure learning;
Inputting the second measured data characteristic and the third simulation data characteristic after the countermeasure learning to a domain discriminator for domain prediction training, and outputting a domain prediction result to construct a second sub-domain self-adaptive model;
And combining the first sub-domain self-adaptive model and the second sub-domain self-adaptive model to construct the domain self-adaptive model.
9. The method for predicting the distance of an underwater target based on the self-adaption of the partially unsupervised field according to claim 8, wherein the constructing a first loss function about the distance prediction training and a second loss function about the domain prediction training, respectively, and weighting the first loss function and the second loss function to obtain a third loss function, comprises:
acquiring a first loss function of the first sub-domain adaptive model;
Calculating a second loss function of the second sub-domain adaptive model according to the following formula:
.
Wherein L D denotes the second loss function, Representing a calculation expectation, wherein P s (c) represents the joint distribution of simulation training samples, P t (x) represents the joint distribution of actual measurement training samples, w represents a weight corresponding to the actual measurement training samples, D represents the generator, G represents the discriminator, c i represents the ith simulation training sample, and x i represents the ith actual measurement training sample;
Weighting the first loss function and the second loss function according to the following formula, thereby obtaining the third loss function:
.
Wherein L represents the third loss function, L C represents the first loss function, A trade-off factor representing the first and second loss functions.
10. The method for predicting the distance between the underwater target based on the self-adaption in the partially unsupervised field according to claim 9, wherein the performing parameter update on the domain self-adaption model by using the third loss function to construct and obtain the underwater target distance prediction model comprises:
And updating parameters of the domain adaptive model according to the following steps:
.
wherein, Model parameters after updating parameters of the domain adaptive model are represented,/>Representing model parameters before parameter updating of the domain adaptive model,/>Representing learning rate,/>Representing the gradient trade-off factor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410319967.XA CN117911852B (en) | 2024-03-20 | 2024-03-20 | Underwater target distance prediction method based on self-adaption in part of unsupervised field |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410319967.XA CN117911852B (en) | 2024-03-20 | 2024-03-20 | Underwater target distance prediction method based on self-adaption in part of unsupervised field |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117911852A true CN117911852A (en) | 2024-04-19 |
| CN117911852B CN117911852B (en) | 2024-06-14 |
Family
ID=90692749
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410319967.XA Active CN117911852B (en) | 2024-03-20 | 2024-03-20 | Underwater target distance prediction method based on self-adaption in part of unsupervised field |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117911852B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112733447A (en) * | 2021-01-07 | 2021-04-30 | 浙江大学 | Underwater sound source positioning method and system based on domain adaptive network |
| CN113392967A (en) * | 2020-03-11 | 2021-09-14 | 富士通株式会社 | Training method of domain confrontation neural network |
| CN116030300A (en) * | 2022-12-29 | 2023-04-28 | 西安电子科技大学 | Progressive domain self-adaptive recognition method for zero-sample SAR target recognition |
| CN117236188A (en) * | 2023-10-12 | 2023-12-15 | 长安大学 | Simulation data-assisted bearing migration diagnosis method, system and equipment medium |
| CN117910346A (en) * | 2024-01-10 | 2024-04-19 | 西北工业大学 | Robust unsupervised field self-adaptive target positioning method based on generalization theory energization |
-
2024
- 2024-03-20 CN CN202410319967.XA patent/CN117911852B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113392967A (en) * | 2020-03-11 | 2021-09-14 | 富士通株式会社 | Training method of domain confrontation neural network |
| CN112733447A (en) * | 2021-01-07 | 2021-04-30 | 浙江大学 | Underwater sound source positioning method and system based on domain adaptive network |
| CN116030300A (en) * | 2022-12-29 | 2023-04-28 | 西安电子科技大学 | Progressive domain self-adaptive recognition method for zero-sample SAR target recognition |
| CN117236188A (en) * | 2023-10-12 | 2023-12-15 | 长安大学 | Simulation data-assisted bearing migration diagnosis method, system and equipment medium |
| CN117910346A (en) * | 2024-01-10 | 2024-04-19 | 西北工业大学 | Robust unsupervised field self-adaptive target positioning method based on generalization theory energization |
Non-Patent Citations (3)
| Title |
|---|
| JINFU FAN: "Partial Label Learning Based on Disambiguation Correction Net With Graph Representation", IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, 31 December 2021 (2021-12-31) * |
| 王彦捷: "改进型VAEGAN的不完整数据模糊聚类算法研究", 中国优秀硕士论文电子期刊网, 15 January 2021 (2021-01-15) * |
| 赵文仓;袁立镇;徐长凯;: "基于鉴别模型和对抗损失的无监督域自适应方法", 高技术通讯, no. 07, 15 July 2020 (2020-07-15) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117911852B (en) | 2024-06-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109993280B (en) | Underwater sound source positioning method based on deep learning | |
| CN109766791B (en) | Communication signal modulation identification method based on self-encoder | |
| Tian et al. | Joint learning model for underwater acoustic target recognition | |
| CN112861066B (en) | Machine learning and FFT (fast Fourier transform) -based blind source separation information source number parallel estimation method | |
| CN113591553B (en) | Turbine pump transfer learning fault intelligent judgment method based on small sample weight optimization | |
| Mao et al. | Automated Bayesian operational modal analysis of the long-span bridge using machine-learning algorithms | |
| CN111695461B (en) | Radar clutter intelligent classification method based on image frequency characteristics | |
| CN114531729A (en) | Positioning method, system, storage medium and device based on channel state information | |
| CN113109794B (en) | Deep sea sound source depth setting method based on deep neural network in strong noise environment | |
| Yang et al. | Underwater acoustic signal prediction based on correlation variational mode decomposition and error compensation | |
| CN116047427A (en) | Small sample radar active interference identification method | |
| Van Komen et al. | A CNN for range and seabed estimation on normalized and extracted time-series impulses | |
| CN113253248B (en) | Small sample vertical array target distance estimation method based on transfer learning | |
| CN118262741A (en) | Ship radiation noise target identification method integrating attention mechanism | |
| Jin et al. | Semi-supervised underwater acoustic source localization based on residual convolutional autoencoder | |
| Li et al. | Hierarchical refined composite multiscale fluctuation-based dispersion entropy: application to feature extraction of underwater target signal | |
| Mohanapriya et al. | Environmental sound recognition using Gaussian mixture model and neural network classifier | |
| CN117911852B (en) | Underwater target distance prediction method based on self-adaption in part of unsupervised field | |
| CN117451055A (en) | Underwater sensor positioning method and system based on basis tracking noise reduction | |
| Guo et al. | Target depth estimation by deep neural network based on acoustic interference structure in deep water | |
| Duangpummet et al. | A robust method for blindly estimating speech transmission index using convolutional neural network with temporal amplitude envelope | |
| CN115586516A (en) | Ship radiation noise identification method based on deep learning and multi-feature extraction | |
| CN114139601A (en) | Evaluation method and system for artificial intelligence algorithm model of power inspection scene | |
| Dehkordi et al. | A compressive sensing based compressed neural network for sound source localization | |
| Khalek et al. | Comparative study of wavelet-SARIMA and wavelet-NNAR models for groundwater level in Rajshahi district |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |