CN116846489B - Optical positioning system and positioning method based on transfer learning - Google Patents
Optical positioning system and positioning method based on transfer learning Download PDFInfo
- Publication number
- CN116846489B CN116846489B CN202310748453.1A CN202310748453A CN116846489B CN 116846489 B CN116846489 B CN 116846489B CN 202310748453 A CN202310748453 A CN 202310748453A CN 116846489 B CN116846489 B CN 116846489B
- Authority
- CN
- China
- Prior art keywords
- node
- optical
- positioning
- coral
- optical receiver
- 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.)
- Active
Links
- 230000003287 optical effect Effects 0.000 title claims abstract description 73
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000013526 transfer learning Methods 0.000 title claims abstract description 22
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 29
- 230000006854 communication Effects 0.000 claims abstract description 23
- 230000006870 function Effects 0.000 claims abstract description 22
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 21
- 238000004891 communication Methods 0.000 claims abstract description 21
- 238000013508 migration Methods 0.000 claims abstract description 12
- 230000005012 migration Effects 0.000 claims abstract description 12
- 235000014653 Carica parviflora Nutrition 0.000 claims description 31
- 241000243321 Cnidaria Species 0.000 claims description 31
- 238000012549 training Methods 0.000 claims description 17
- 239000013598 vector Substances 0.000 claims description 13
- 238000000605 extraction Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000013507 mapping Methods 0.000 claims description 6
- 230000007246 mechanism Effects 0.000 claims description 6
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000006978 adaptation Effects 0.000 claims description 3
- 238000004088 simulation Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 5
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 5
- 230000005540 biological transmission Effects 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06N3/096—Transfer learning
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B13/00—Transmission systems characterised by the medium used for transmission, not provided for in groups H04B3/00 - H04B11/00
- H04B13/02—Transmission systems in which the medium consists of the earth or a large mass of water thereon, e.g. earth telegraphy
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/20—Monitoring; Testing of receivers
- H04B17/27—Monitoring; Testing of receivers for locating or positioning the transmitter
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/309—Measuring or estimating channel quality parameters
- H04B17/318—Received signal strength
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/391—Modelling the propagation channel
Abstract
The invention discloses an optical positioning system based on transfer learning and a node position estimation method thereof, wherein the optical system comprises a plurality of communication nodes, and each communication node comprises a transmitting node and a receiving node; the receiving node comprises a plurality of optical receivers, wherein the plurality of optical receivers are distributed in a circular multi-layer mode, and the number of the optical receivers from the inner layer to the outer layer is multiplied. The position estimation method comprises the steps of establishing an optical receiver coordinate system of a receiving node, calculating an RSS value of each optical receiver, establishing a convolutional neural network positioning model according to the obtained RSS value, establishing a migration learning algorithm architecture, counting a loss function through the established migration learning algorithm architecture, updating model parameters through a BP algorithm and an optimizer to minimize the loss function, and carrying out accurate positioning. The invention overcomes the problems of poor expandability, poor robustness and instantaneity, high cost and the like of a centralized network positioning scheme by providing a one-to-multiple-receiving underwater laser communication scene and a node structure.
Description
Technical Field
The invention belongs to the technical field of underwater sensing, and particularly relates to an optical positioning system based on transfer learning and a node position estimation method thereof.
Background
An Underwater Wireless Sensor Network (UWSN) is one of the indispensable technologies in modern marine monitoring systems, and can provide a large amount of data and information for the fields of marine science, environmental protection, military safety, and the like. In the UWSN, the wireless communication technology is one of the keys, because physical characteristics of the underwater environment impose serious restrictions on communication transmission. Currently, underwater Wireless Optical Communication (UWOC) is widely used in UWSN as a fast, inexpensive, low-power communication technology to increase transmission rate and reduce energy consumption. However, there are some challenges and limitations in UWOC communication, the most significant of which is laser alignment. Laser alignment presents a number of challenges due to the dynamics and complexity of the underwater environment. For example, disturbances in the flow of water, obstruction of underwater objects, instability of the underwater environment, etc., all have an impact on the stability of the communication link. Thus, laser alignment is one of the key issues that need to be addressed in UWOC. To ensure the stability of the communication link, the transmitting node needs to precisely position the receiving node for alignment before the communication link is established. The accuracy and reliability of positioning is critical to the success of UWOC communications.
In the current practical application, the UWOC positioning mainly adopts a centralized network positioning scheme. Such centralized network positioning schemes typically require the deployment of additional positioning nodes under water, which enable the estimation and positioning of the node's position by receiving signals from the nodes and calculating parameters such as time differences. Although the method can realize higher positioning precision, additional hardware nodes are required to be deployed, so that the complexity and the cost of the system are increased, and the expandability of the system is limited. When the network scale is large, the number of nodes is excessive, so that the nodes are mutually interfered, and the performance and the positioning accuracy of the system are affected. In addition, centralized network positioning schemes are also affected by the underwater environment. The underwater environment is complex and changeable, and the positioning accuracy of the nodes can be affected by factors such as water flow, water temperature and the like. For centralized network positioning schemes, additional positioning nodes need to be added to the network, and deployment and operation of the nodes are also affected by the underwater environment, thereby affecting the performance and stability of the system.
Disclosure of Invention
The invention aims to provide an optical positioning system based on transfer learning and a node position estimation method thereof, which are applied to an underwater environment by providing a multi-shot laser communication scene and a node structure, so as to solve the problems of poor expandability, poor robustness and real-time performance, high cost and the like of a centralized network positioning scheme.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
an optical positioning system based on transfer learning comprises a plurality of communication nodes, wherein each communication node comprises a transmitting node and a receiving node; the receiving node comprises a plurality of optical receivers, wherein the plurality of optical receivers are distributed in a circular multi-layer mode, and the number of the optical receivers from the inner layer to the outer layer is multiplied.
Further, the transmitting node comprises a transmitting laser source, the adjacent layers of the optical receivers are equidistant, and each optical receiver receives light intensity signals with different intensities sent by the transmitting laser source respectively.
The invention also discloses a positioning method based on the underwater node optical system, which reduces the problem of positioning performance reduction caused by field deviation through a field self-adaptive transfer learning method. In the training process, the inter-domain feature difference of the feature extraction module and the prediction loss of the network model are minimized to improve the positioning performance under different scenes.
The specific technical scheme is as follows:
a node position estimation method based on transfer learning comprises the following steps:
A. establishing an optical receiver coordinate system of a receiving node, simulating a relative position relation between the optical receiver coordinate system and a laser light source of a transmitting node, and calculating an RSS value of each optical receiver;
B. building a convolutional neural network positioning model based on the RSS value obtained in the step A;
C. based on the established convolutional neural network positioning model, a transfer learning algorithm architecture is established, a loss function is counted through the established transfer learning algorithm architecture, model parameters are updated through a BP algorithm and an optimizer to minimize the loss function, and accurate positioning is performed.
Further, in the step a, the optical receiver coordinate system establishes a rectangular coordinate system with the central receiver of the optical receiver group as the origin O, and each optical receiver is a rectangular coordinate systemThe coordinates K of the receiver i (x i ,y i ,z i ) The calculation is shown in the first figure,
further, in step A, the average light power P of the laser light source is randomly generated when the relative positional relationship simulation is performed t Randomly generating an angle alpha, an angle beta, which can uniquely determine the direction of the light beam, and a linear distance L' of the laser source from the established optical receiver coordinate system along the direction of the light beam, the sitting mark of the laser source projection point P is (x P ,y P 0), the position calculation of the laser light source T is shown as formula two,
further, according to the position of the obtained laser light source T, the propagation radius r of each optical receiver perpendicular to the direction of the emission node i And distance L 'from the laser transmitter' i The method comprises the following steps of:
and repeatedly calculates the RSS value of each optical receiver,
R(t)＝ρP t h pl x(t)+w(t)＝ρB(r,L')S r x (t) +w (t), formula four, wherein B (R, L') is the light intensity received by the optical receiver, and generates a database from the generated RSS values, the database is characterized by the RSS sequence of the receiver set as a feature vector r= [ R ] 1 ,R 2 ,…,R N ]The three-dimensional coordinates and pitch angle of the laser source are used as label vectors, wherein N s Representing the number of data samples in the location fingerprint database.
Further, the convolutional neural network positioning model established in the step B comprises an input module, an attention module, a feature extraction module and an output module; the input module inputs as feature vectors into a convolutional neural network positioning model framework through SNR sequences of a plurality of optical receivers on nodes.
Further, the attention module is based on an attention mechanism introduced in the CNN model, and enables the convolutional neural network positioning model to autonomously learn and acquire importance degrees of RSS data of a plurality of channels on positioning through adding an SE module in SENet, and then data useful for network performance is improved based on the importance degrees, and invalid data is restrained.
In step C, in the built migration learning algorithm architecture, feature processing is performed on the source domain data and the target domain data input by the convolutional neural network positioning model, and the MMD loss L between the domains is calculated MMD And CORAL loss L CORAL The MSE loss between the source domain output and the true value is recalculated.
Further, the MMD penalty is shown in equation five,
wherein,
a sixth formula is deep migratable features extracted from the source domain data;
seventh, the deep migratable features extracted from the target domain data are obtained; />And->Mapping functions representing the characteristics of the source domain and the target domain, respectively, in the regenerated kernel Hilbert space>The transformed mean value.
Further, the CORAL loss L CORAL As shown in the eighth view,
wherein C is s And C t Covariance matrices of source domain data and target domain data features, respectively, can be calculated by the following formula nine:
wherein 1 is T A row vector with m elements of all 1 is represented.
Further, after constraining the mean and covariance between the extracted features of the source and target domains, the migration learning algorithm architecture works by minimizing the loss function as shown in the following equation ten:
L＝L s +λ 1 L MMD +λ 2 L CORAL a formula ten;
where L is the loss function when the target domain training dataset is completely untagged, lambda 1 And lambda (lambda) 2 Weight coefficients, L, representing MMD and CORAL differences, respectively s MSE loss representing source domain network predicted and true values:
wherein N is s Represents the training data sample size, phi i Is the true parameter value to be estimated for the ith data sample,the predicted value of the source domain is obtained after predictive regression.
Further, the MMD penalty L at each iteration is obtained separately MMD And CORAL loss L CORAL And MSE loss L s Then, the weight coefficient mu of the MMD and CORAL difference of the current iteration is respectively obtained through self-adaptive adjustment 1 Sum mu 2 Wherein, the method comprises the steps of, wherein,
formula twelve; />Formula thirteen;
and iterating for several times, and obtaining the weight coefficient lambda of the difference between MMD and CORAL of the current training by an averaging method 1 And lambda (lambda) 2 The influence on model training is dynamically balanced, and the domain adaptation learning capacity of the convolutional neural network positioning model is improved.
Compared with the prior art, the invention has the advantages that: firstly, the invention provides a distributed underwater optical positioning and network model based on position fingerprint positioning, and simultaneously, a receiving end obtains the three-dimensional position of a laser source and the direction of a light beam, so that the link alignment in the subsequent communication process is realized to achieve high-quality information transmission.
Secondly, due to complexity and diversity of the underwater environment, the invention completes positioning tasks under different scenes through the built depth self-adaptive convolutional neural network model. Compared with the traditional centralized network positioning scheme, the invention does not need to additionally add positioning nodes, thereby improving the expandability of the system and reducing the cost of the system; the attention-introducing mechanism SE module helps the model to better capture key information in input received signal strength data so as to improve the positioning performance of the network model; and integrating various field self-adaptive methods in the network model to realize how to migrate the learning ability of the previous task to the positioning scheme under the new task under different positioning scenes.
At a training data volume of 2×10 5 When the emitted light power is 10dBm, the average positioning error of the scheme is 0.79m, the MSE is 1.04m, and the method is similar to the prior artThe average positioning error of the technical index is larger than 4m, and the MSE is larger than 6m, so that the method is greatly improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present invention and should not be considered limiting the scope, and that other related drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic structural diagram of an optical positioning system based on transfer learning provided by the invention.
Fig. 2 is a schematic structural diagram of a receiving node of the optical positioning system based on transfer learning provided by the invention.
Fig. 3 is a diagram of a database architecture provided by the present invention.
Fig. 4 is a framework diagram of a migration-based learning algorithm architecture.
Fig. 5 is a diagram of an RSS-based underwater node positioning CNN structure.
Fig. 6 is a graph of experimental data using the present invention.
FIG. 7 is a graph II of experimental data using the present invention.
Reference numerals: 1. a transmitting node; 2. the receiving node.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present 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 noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further definition or explanation thereof is necessary in the following figures.
In the description of the present invention, it should be noted that, if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. indicate an azimuth or a positional relationship based on that shown in the drawings, or an azimuth or a positional relationship in which the inventive product is conventionally put in use, it is merely for convenience of describing the present invention and simplifying the description, and it is not indicated or implied that the apparatus or element referred to must have a specific azimuth, be constructed and operated in a specific azimuth, and thus should not be construed as limiting the present invention.
Furthermore, the terms "first," "second," "third," and the like, if any, are used merely for distinguishing between descriptions and not for indicating or implying a relative importance.
Furthermore, the terms "horizontal," "vertical," "overhang" and the like, if any, do not denote a requirement that the component be absolutely horizontal or overhang, but rather may be slightly inclined. As "horizontal" merely means that its direction is more horizontal than "vertical", and does not mean that the structure must be perfectly horizontal, but may be slightly inclined.
In the description of the present invention, it should also be noted that, unless explicitly stated and limited otherwise, the terms "disposed," "mounted," "connected," and the like are to be construed broadly, and may be, for example, fixedly connected, detachably connected, or integrally connected; can be mechanically or electrically connected; can be directly connected or indirectly connected through an intermediate medium, and can be communication between two elements. The specific meaning of the above terms in the present invention will be understood in specific cases by those of ordinary skill in the art.
It should be noted that the features of the embodiments of the present invention may be combined with each other without conflict.
As shown in fig. 1-5, an optical positioning system based on transfer learning includes a plurality of communication nodes, and each communication node includes a transmitting node 1 and a receiving node 2; the receiving node 2 comprises a plurality of optical receivers, the plurality of optical receivers are distributed in a circular multi-layer mode, and the number of the optical receivers from the inner layer to the outer layer is multiplied; the transmitting node 1 comprises a transmitting laser source, the adjacent layers of the optical receivers are equidistant, and each optical receiver respectively receives light intensity signals with different intensities emitted by the transmitting laser source.
The invention also discloses a positioning method based on the underwater node optical system, which comprises the following specific technical scheme:
a node position estimation method based on transfer learning comprises the following steps:
A. establishing an optical receiver coordinate system of a receiving node, simulating a relative position relation between the optical receiver coordinate system and a laser light source of a transmitting node, and calculating an RSS value of each optical receiver;
B. building a convolutional neural network positioning model based on the RSS value obtained in the step A;
C. based on the established convolutional neural network positioning model, a transfer learning algorithm architecture is established, a loss function is counted through the established transfer learning algorithm architecture, model parameters are updated through a BP algorithm and an optimizer to minimize the loss function, and accurate positioning is performed.
Compared with the prior art, the invention provides and designs a one-to-many-to-one underwater laser communication scene and a node structure, and a transmitting laser and a plurality of receivers are distributed on the node. The relative positional relationship of the transmitting node and the receiving node is modeled, the signal strength (RSS) of the receiver is calculated using the beam propagation function and a database of the received signal strengths of the plurality of receivers versus the position of the transmitting laser and the direction of the beam is built.
Aiming at the problems of interference, great influence by underwater environment, poor positioning precision and the like caused by excessive number of positioning nodes of a centralized network, an underwater optical convolutional neural network (UO-CNN) positioning model is provided, and a concentration mechanism SE module is introduced into the convolutional neural network to improve the positioning performance of the system.
Aiming at the variability of underwater positioning scenes and the difficulty of data acquisition, the invention utilizes a field self-adaption based migration learning method to reduce the problem of positioning performance reduction caused by field deviation. The first order statistic difference and the second order statistic difference between the source domain and the target domain features after the feature extraction module are measured by using a maximum mean value difference (MMD) and a CORAL function respectively, and the inter-domain feature difference of the feature extraction module and the prediction loss of a network model are minimized in the training process to improve the positioning performance under different scenes.
In the method for optical positioning of underwater nodes based on transfer learning provided by the invention, the construction of the nodes is needed first, in the step A, the coordinate system of each optical receiver is established by using the central receiver of the optical receiver group as the origin O in the optical receiver coordinate system, and the coordinate K of each optical receiver is established in the rectangular coordinate system i (x i ,y i ,z i ) The calculation is shown in the first figure,
in step A, the average light power P of the laser light source is randomly generated when the relative position relation simulation is performed t The angle alpha, angle beta, and the linear distance L' of the laser source along the beam direction to the established optical receiver coordinate system, which uniquely determine the direction of the beam, are randomly generated. The angle alpha is the included angle between the laser beam and the XOY plane, alpha is 0,0.5 pi, the angle beta is the included angle between the projection of the laser beam on the XOY plane and the X axis, and the angle beta is 0,2 pi]。
The coordinates of the projection point P of the laser light at XOY are randomly generated within a certain range. After the water quality attenuation coefficient c is determined, the maximum detectable shift radius r of the laser beam can be determined according to the laser scanning range L m . Obtaining L' and r from multiple sets of values m After the dispersion curve of (2), a continuous (L' -r) is obtained by polynomial fitting m ) A curve. Based on the fitted (L' -r) m ) Curve, calculating maximum detectable offset radius r corresponding to randomly generated value L m In order to make the receiver on the node detect effective data as much as possible, the coordinates of the projection point P of the laser source at XOY are taken as (x P ,y P 0), wherein x P And y P Are all within the range of [ -0.8r m ,0.8r m ]Random numbers within.
The sitting according to the laser source projection point P is marked as (x) P ,y P 0), and randomly generated angles alpha, beta, the position (x, y, z) of the laser light source T is calculated as shown in equation two,
according to the position of the obtained laser light source T, the propagation radius r of each optical receiver perpendicular to the direction of the emission node i And distance L 'from the laser transmitter' i The method comprises the following steps of:
and repeatedly calculates the RSS value of each optical receiver,
R(t)＝ρP t h p lx(t)+t(t)＝ρB(r,L')S r x (t) +w (t), formula four,
where B (r, L') is the intensity of light received by the optical receiver,
repeating the calculation process, calculating the RSS value of each receiver, generating a database by using the generated RSS value, wherein the database is characterized by using the RSS sequence of the receiver group as a characteristic vector R= [ R ] 1 ,R 2 ,…,R N ]The three-dimensional coordinates and pitch angle of the laser source are used as label vectors, wherein N s Representing the number of data samples in the location fingerprint database.
The convolutional neural network positioning model established in the step B comprises an input module, an attention module, a feature extraction module and an output module, wherein:
the input module is used for inputting the SNR sequences of the plurality of optical receivers on the nodes into the convolutional neural network positioning model framework as the characteristic vectors.
The attention module, the network can learn and acquire the importance degree of the RSS data of N channels on different receivers on the positioning autonomously, then based on the importance degree, the data useful for the network performance is improved, and the data with little use is restrained. The data output from this module is the original received vector r= [ R ] for each receiver in the channel dimension 1 ,R 2 ,…,R N ]The data after recalibration of the importance degree of the RSS data of each receiver, namelyBy utilizing the module based on the channel attention mechanism and combining with the strong feature extraction and mapping capability of CNN, the positioning performance of the system is integrally improved.
The feature extraction module comprises four hidden layers in total, wherein each hidden layer consists of a convolution layer, a ReLU layer and a normalization layer (Batch Normalization, BN), and the dimension of each convolution layer is 64 multiplied by 15. Features of recalibration data output from the SENet module are extracted through convolution operation of a convolution layer, nonlinear mapping is conducted on a convolution result through a ReLU activation function, gradient disappearance and gradient explosion problems in the BP process are avoided, and convergence speed of a model is improved. BN operations are then used to reduce the variance between samples and to speed up the convergence speed of the network.
The output module comprises a flat layer, wherein the flat layer unidimensionally outputs the upper layer, then the flat layer is connected with the full-connection layer, and the three-dimensional coordinates (x, y, z) of the laser light source to be positioned, the angle alpha and the angle beta of the light beam are predicted and output.
After the whole positioning system is introduced into the attention module, the model is increased by fewer parameters, but the more accurate nonlinear relation between the RSS data of the receiver for positioning and the position and the beam direction of the laser source is established by adding the strong mapping capability of the CNN network, so that the positioning model with better performance is obtained.
In the step C, in the built migration learning algorithm architecture, the overall hierarchy architecture and each layer of parameter setting of the network model are consistent with the CNN positioning model. In a different way, the difference is that,in the training stage, the network inputs not only source domain data with large data volume and labels, but also a small amount of target domain data, and all the target domain data are not labeled. After the data of the source domain and the target domain pass through the feature extraction module, the two-dimensional features are flattened into one dimension through the flat layer, and then MMD loss L between the domains is calculated through a fully-connected self-adaptive layer MMD And CORAL loss L CORAL . Finally, after the prediction output of the last full-connection layer is carried out, the MSE loss between the source domain output and the true value is calculated. The loss function of the whole transfer learning network model is the superposition of characteristic difference loss between fields and MSE loss of network output and a true value, model parameters are updated by minimizing the loss function through a BP algorithm and an optimizer, so that the field characteristic gap is reduced, the target field positioning performance is improved, and the result of the transfer learning algorithm architecture positioning stability is enhanced.
While the MMD loss is shown in equation five,
wherein,
a sixth formula is deep migratable features extracted from the source domain data;
seventh, the deep migratable features extracted from the target domain data are obtained; />And->Mapping functions representing the characteristics of the source domain and the target domain, respectively, in the regenerated kernel Hilbert space>The transformed mean value.
While CORAL loses L CORAL As shown in the eighth view,
wherein C is s And C t Covariance matrices of source domain data and target domain data features, respectively, can be calculated by the following formula nine:
wherein 1 is T A row vector with m elements of all 1 is represented.
In summary, after constraining the mean and covariance between the extracted features of the source and target domains, the migration learning algorithm architecture works by minimizing the loss function as shown in the following equation ten:
L＝L s +λ 1 L MMD +λ 2 L CORAL a formula ten;
where L is the loss function when the target domain training dataset is completely untagged, lambda 1 And lambda (lambda) 2 Weight coefficients, L, representing MMD and CORAL differences, respectively s MSE loss representing source domain network predicted and true values:
wherein N is s Represents the training data sample size, phi i Is the true parameter value to be estimated for the ith data sample,the predicted value of the source domain is obtained after predictive regression.
MMD loss L at each respective iteration MMD And CORAL lossLoss of L CORAL And MSE loss L s Then, the weight coefficient mu of the MMD and CORAL difference of the current iteration is respectively obtained through self-adaptive adjustment 1 Sum mu 2 Wherein, the method comprises the steps of, wherein,
formula twelve; />Formula thirteen;
and iterating for several times, and obtaining the weight coefficient lambda of the difference between MMD and CORAL of the current training by an averaging method 1 And lambda (lambda) 2 The influence on model training is dynamically balanced, and the domain adaptation learning capacity of the convolutional neural network positioning model is improved.
In summary, according to the optical positioning system and the positioning method thereof based on transfer learning provided by the invention, the distributed underwater optical positioning system is provided, and the receiving end obtains the three-dimensional position of the laser source and the direction of the light beam, so that the link alignment in the subsequent communication process is realized to achieve high-quality information transmission. Due to complexity and diversity of the underwater environment, a depth self-adaptive network model is provided for completing positioning tasks under different scenes. Compared with the traditional centralized network positioning scheme, the invention does not need to additionally add positioning nodes, thereby improving the expandability of the system and reducing the cost of the system; the attention-introducing mechanism SE module helps the model to better capture key information in input received signal strength data so as to improve the positioning performance of the model; the model is fused with a plurality of field self-adaptive methods to realize how to transfer the learning ability of the previous task to the positioning scheme under the new task under different positioning scenes, thereby realizing more excellent positioning.
In experimental practice, the training data amount is 2×10 5 And when the emitted light power is 10dBm, as shown in fig. 6 and 7, the average positioning error of the scheme is 0.79m, the MSE is 0.87, and compared with the average positioning error of the prior art index which is greater than 4m and the MSE which is greater than 6, the average positioning error of the prior art index is greatly improved.
The above embodiments are merely for illustrating the technical solution of the present application, and not for limiting the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.
Claims (2)
1. The node position estimation method based on the transfer learning is characterized by comprising the following steps of:
A. establishing an optical receiver coordinate system of a receiving node, simulating a relative position relation between the optical receiver coordinate system and a laser light source of a transmitting node, and calculating an RSS value of each optical receiver;
B. building a convolutional neural network positioning model based on the RSS value obtained in the step A;
C. based on the established convolutional neural network positioning model, establishing a migration learning algorithm architecture, counting a loss function through the established migration learning algorithm architecture, updating model parameters through a BP algorithm and an optimizer minimum loss function, and performing accurate positioning;
in the step A, the optical receiver coordinate system establishes a rectangular coordinate system by taking the central receiver of the optical receiver group as the origin O, and the coordinate K of each optical receiver i (x i ,y i ,z i ) The calculation is shown in the first figure,
in step A, the average light power P of the laser light source is randomly generated when the relative position relation simulation is performed t Randomly generating an angle alpha, an angle beta which can uniquely determine the direction of a light beam and a linear distance L' of a laser source from an established optical receiver coordinate system along the direction of the light beam, wherein the sitting mark of a projection point P of the laser source is as follows(x P ,y P 0), the position calculation of the laser light source T is shown as formula two,
according to the position of the obtained laser light source T, the propagation radius r of each optical receiver perpendicular to the direction of the emission node i And distance L 'from the laser transmitter' i The method comprises the following steps of:
and repeatedly calculates the RSS value of each optical receiver,
R(t)＝ρP t h pl x(t)+w(t)＝ρB(r,L')S r x (t) +w (t), formula four, wherein B (R, L') is the light intensity received by the optical receiver, and generates a database from the generated RSS values, the database is characterized by the RSS sequence of the receiver set as a feature vector r= [ R ] 1 ,R 2 ,…,R N ]The three-dimensional coordinates and pitch angle of the laser source are used as label vectors, wherein N s Representing the number of data samples in the location fingerprint database;
the convolutional neural network positioning model established in the step B comprises an input module, an attention module, a feature extraction module and an output module; the input module is used for inputting the SNR sequences of a plurality of optical receivers on the nodes into a convolutional neural network positioning model frame as feature vectors; the attention module is based on an attention mechanism introduced into a CNN model, and enables a convolutional neural network positioning model to autonomously learn and acquire importance degrees of RSS data of a plurality of channels on positioning by adding an SE module in SENet, and then, based on the importance degrees, data useful for network performance are improved and invalid data are restrained;
in the step C, in the built migration learning algorithm architecture, the characteristic processing is carried out on the source domain data and the target domain data input by the convolutional neural network positioning model, and the calculation is carried outCalculating MMD loss L between fields MMD And CORAL loss L CORAL Calculating MSE loss between the source domain output and the true value;
the MMD loss is shown in equation five,
wherein,
deep migratable features extracted for source domain data;
deep migratable features extracted for the target domain data; />And->Mapping functions representing the characteristics of the source domain and the target domain, respectively, in the regenerated kernel Hilbert space H>The transformed mean value;
the CORAL loss L CORAL As shown in the eighth view,
wherein C is s And C t Covariance matrices of source domain data and target domain data features, respectively, can be calculated by the following formula nine:
wherein 1 is T Representing a row vector with m elements of all 1;
after constraining the mean and covariance between the extracted features of the source and target domains, the migration learning algorithm architecture works by minimizing the loss function as shown in the following equation ten:
L＝L s +λ 1 L MMD +λ 2 L CORAL a formula ten;
where L is the loss function when the target domain training dataset is completely untagged, lambda 1 And lambda (lambda) 2 Weight coefficients, L, representing MMD and CORAL differences, respectively s MSE loss representing source domain network predicted and true values:
wherein N is s Represents the training data sample size, phi i Is the true parameter value to be estimated for the ith data sample,obtaining a predicted value of a source domain after predictive regression;
MMD loss L at each respective iteration MMD And CORAL loss L CORAL And MSE loss L s Then, the weight coefficient mu of the MMD and CORAL difference of the current iteration is respectively obtained through self-adaptive adjustment 1 Sum mu 2 Wherein, the method comprises the steps of, wherein,
and iterating for several times, and obtaining the weight coefficient lambda of the difference between the MMD and the CORAL of the current training by an averaging method 1 And lambda (lambda) 2 Come fromThe influence of dynamic balance on model training improves the domain adaptation learning capacity of a convolutional neural network positioning model.
2. An optical positioning system based on the transfer learning node position estimation method according to claim 1, characterized by comprising a plurality of communication nodes, and each communication node comprising a transmitting node (1) and a receiving node (2); the receiving node (2) comprises a plurality of optical receivers, wherein the plurality of optical receivers are distributed in a circular multi-layer mode, and the number of the optical receivers from the inner layer to the outer layer is increased in multiple; the transmitting node (1) comprises a transmitting laser source, the adjacent layers of the optical receivers are equidistant in interval, and each optical receiver receives light intensity signals with different intensities emitted by the transmitting laser source respectively.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310748453.1A CN116846489B (en) | 2023-06-25 | 2023-06-25 | Optical positioning system and positioning method based on transfer learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310748453.1A CN116846489B (en) | 2023-06-25 | 2023-06-25 | Optical positioning system and positioning method based on transfer learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116846489A CN116846489A (en) | 2023-10-03 |
| CN116846489B true CN116846489B (en) | 2024-01-30 |
Family
ID=88171845
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310748453.1A Active CN116846489B (en) | 2023-06-25 | 2023-06-25 | Optical positioning system and positioning method based on transfer learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116846489B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109117793A (en) * | 2018-08-16 | 2019-01-01 | 厦门大学 | Direct-push high Resolution Range Profile Identification of Radar method based on depth migration study |
| CN116028876A (en) * | 2022-09-20 | 2023-04-28 | 北京工业大学 | Rolling bearing fault diagnosis method based on transfer learning |
| US11658752B1 (en) * | 2022-01-21 | 2023-05-23 | Qinghai Normal University | Node positioning method for underwater wireless sensor network (UWSN) based on zeroing neural dynamics (ZND) |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108387867A (en) * | 2018-02-09 | 2018-08-10 | 华南理工大学 | A kind of underwater source node localization method |
-
2023
- 2023-06-25 CN CN202310748453.1A patent/CN116846489B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109117793A (en) * | 2018-08-16 | 2019-01-01 | 厦门大学 | Direct-push high Resolution Range Profile Identification of Radar method based on depth migration study |
| US11658752B1 (en) * | 2022-01-21 | 2023-05-23 | Qinghai Normal University | Node positioning method for underwater wireless sensor network (UWSN) based on zeroing neural dynamics (ZND) |
| CN116028876A (en) * | 2022-09-20 | 2023-04-28 | 北京工业大学 | Rolling bearing fault diagnosis method based on transfer learning |
Non-Patent Citations (1)
| Title |
|---|
| 基于深度学习的水下光学图像分类识别方法研究;汤永涛等;舰船电子工程(第8期);189-194 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116846489A (en) | 2023-10-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106093849B (en) | A kind of Underwater Navigation method based on ranging and neural network algorithm | |
| CN108051779B (en) | TDOA-oriented positioning node optimization method | |
| CN111901802A (en) | MISO system downlink secrecy rate optimization method by means of intelligent reflection surface | |
| CN111323752B (en) | Far and near field transition interval sound source positioning method | |
| CN111200456B (en) | Fast and low-consumption 3D beam forming method based on joint autonomous positioning | |
| CN111818449A (en) | Visible light indoor positioning method based on improved artificial neural network | |
| CN112135344B (en) | CSI (channel State information) and DCNN (distributed channel neural network) -based equipment-free target positioning method | |
| CN112468230A (en) | Wireless ultraviolet light scattering channel estimation method based on deep learning | |
| CN112612001A (en) | Track prediction method and device based on BP neural network algorithm | |
| CN112986907A (en) | Moving target positioning method under clock deviation and clock drift conditions | |
| CN116125386A (en) | Intelligent positioning method and system for underwater vehicle with enhanced sparse underwater acoustic ranging | |
| CN116846489B (en) | Optical positioning system and positioning method based on transfer learning | |
| CN108759846B (en) | Method for establishing self-adaptive extended Kalman filtering noise model | |
| CN112346010B (en) | Dual-computer passive positioning method based on scale difference and time difference | |
| CN107483131B (en) | Method for generating high-speed aircraft double-satellite combined channel Markov state sequence | |
| CN111505566A (en) | Ultrahigh frequency radio frequency signal DOA estimation method | |
| CN105372644B (en) | One kind is based on the modified Adaptive beamformer method and system of dynamic weight | |
| CN111812580A (en) | Motion linear sparse array optimization method based on underdetermined information source Cramer-Rao bound | |
| CN108924734B (en) | Three-dimensional sensor node positioning method and system | |
| Wu et al. | Research on RSS based indoor location method | |
| CN114167347B (en) | Amplitude-phase error correction and direction finding method for mutual mass array in impact noise environment | |
| Ming et al. | Study on the personnel localization algorithm of the underground mine based on rssi technology | |
| CN112954637B (en) | Target positioning method under condition of uncertain anchor node position | |
| CN115616477A (en) | Non-plane wave monopulse angle measurement method | |
| CN114415157A (en) | Underwater target multi-model tracking method based on underwater acoustic sensor network |
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 |