A fingerprint database positioning method based on CWGAN and neural network

Abstract

The application provides a fingerprint database positioning method based on a CWGAN and a neural network, which is applied to a 5G indoor scene and comprises the following steps: S1, a first positioning model is obtained through training, including the following steps: S11, an offline fingerprint database is acquired and filtered to obtain a filtered offline fingerprint database, and the filtered offline fingerprint database is normalized to obtain a first offline fingerprint database; S12, the first offline fingerprint database is processed based on a CWGAN to generate a new fingerprint, and the new fingerprint constitutes a second offline fingerprint database; S13, the first offline fingerprint database and the second offline fingerprint database are combined to obtain a combined offline fingerprint database; S14, a CNN-LSTM model is acquired, the CNN-LSTM model is trained, and the first positioning model is obtained; S2, the fingerprints of test points are collected and processed according to step S1; and S3, the processed fingerprints of the test points are input into the first positioning model, and the first positioning model outputs predicted coordinate values of the test points. The fingerprint database positioning method based on the CWGAN and the neural network has small positioning errors and high positioning precision.

Description

Technical Field

[0001] This invention belongs to the field of positioning and navigation, and particularly relates to a fingerprint database positioning method based on CWGAN and neural networks. Background Technology

[0002] Currently, the development of 5G enables it to support higher speed and lower latency positioning, but high data collection costs and fluctuations in indoor signal reception remain the main factors restricting efficient and high-precision positioning in indoor scenarios using 5G.

[0003] Indoor environments often contain obstacles such as walls, floors, ceilings, and metal shielding, which cause signal reflection and scattering. This multipath propagation exhibits non-stationary characteristics in both time and space, leading to decreased indoor positioning accuracy and fluctuations in positioning results over time. While deep learning-based fingerprint matching algorithms have been widely used for indoor positioning, training these models requires vast amounts of data, and data collection is extremely costly in terms of manpower and time, significantly impacting the efficiency of 5G indoor positioning.

[0004] Based on the above, this application provides a technical solution to solve the above technical problems. Summary of the Invention

[0005] To address the issues of low accuracy, large fluctuations, and low efficiency in existing 5G indoor positioning technologies, this invention provides a fingerprint database-based positioning method based on CWGAN and neural networks, applicable to 5G indoor scenarios, including:

[0006] Step S1: Train the first localization model, including:

[0007] Step S11: Obtain and filter the offline fingerprint database to obtain the filtered offline fingerprint database. Normalize the filtered offline fingerprint database to obtain the first offline fingerprint database. The fingerprints constituting the offline fingerprint database are the SS-RSRP values ​​of the synchronization signal reference received power of the 5G base station primary and neighboring cells collected by the device in the positioning area.

[0008] Step S12: Process the first offline fingerprint database based on CWGAN. The CWGAN learns the feature distribution of the fingerprints in the first offline fingerprint database and generates a new fingerprint that is similar to the feature distribution of the fingerprints in the first offline fingerprint database. The generated new fingerprint constitutes the second offline fingerprint database.

[0009] Step S13: Merge the first offline fingerprint database and the second offline fingerprint database to obtain a merged offline fingerprint database;

[0010] Step S14: Obtain the CNN-LSTM model. Input the merged offline fingerprint database into the CNN-LSTM model. When the parameters of the CNN-LSTM model meet the preset conditions, complete the training of the CNN-LSTM model and obtain the first localization model.

[0011] Step S2: Collect fingerprints at the test points and process the data according to step S1;

[0012] Step S3: Input the processed fingerprint of the test point into the first positioning model, and the first positioning model outputs the predicted coordinate value of the test point.

[0013] Preferably, obtaining the offline fingerprint database in step S11 includes:

[0014] Step S111: Divide the positioning area into N The number of grids and the number of 5G base stations within the positioning area are: M Each grid intersection is designated as a reference point. The fingerprint of each reference point is obtained, and the fingerprint feature vectors of all reference points constitute an offline fingerprint database. , ;

[0015] in, It is the first i A reference point at a certain time t The fingerprint formed by the SS-RSRPs of all 5G base stations received internally, 1≤ i≤N .

[0016] Preferably, the filtering of the offline fingerprint database in step S11 includes:

[0017] Step S112, the i The feature vector of a fingerprint from a single sample at each reference point is represented as follows: ,delete Vectors containing null values ​​are selected, and those with SS-RSRP values ​​higher than a threshold are filtered out. The sample is used as the fingerprint of this reference point. ;

[0018] Step S113: Obtain the time t The extreme points of the received fingerprint are identified, and the upper envelope is obtained by curve interpolation based on these extreme points. and lower envelope and the upper envelope and the lower envelope Calculate the average to obtain the average envelope. And calculate the remaining components. :

[0019] ;

[0020] Step S114, Repeat step S113 as a new signal until the difference between the number of extreme points and zero-crossing points is no greater than 1, and the average value of the envelope of the maximum and minimum values ​​at any point is 0. At this point, the remaining components... Components of the first-order eigenmode function ;

[0021] Step S115: Calculate the first-order residual: ;

[0022] make Alternative Repeat steps S113 to S114 until the residual is monotonic or no longer oscillates.

[0023] Step S116, Discard the previous steps p The first-order intrinsic mode function component is obtained by summing all the remaining intrinsic mode function components to obtain the filtered fingerprint of the reference point. n is the total number of intrinsic mode function components;

[0024] Step S117: Merge all reference points Obtain the filtered offline fingerprint database .

[0025] Preferably, step S11 further includes:

[0026] Step S118: Filter the offline fingerprint database Normalize each column, the first... k The normalized SS-RSRP value is expressed as follows: , Indicates the first k The minimum value of SS-RSRP. Indicates the first k The maximum value of SS-RSRP in each column is used to merge the results of each column to form the normalized first offline fingerprint database. , normed This indicates normalization processing;

[0027] Step S119: Use one-hot encoding to process the numbering of each reference point. One-hot codes for all corresponding reference points: ,in, For the first i One-hot codes corresponding to each sample The total number of samples.

[0028] Preferably, CWGAN includes a generator and a discriminator, which are respectively composed of a multilayer perceptron (MLP) and a fully connected (FC) layer. Step S12 includes:

[0029] Step S121: Take the one-hot code C of all reference points as a priori conditions, and combine it with the random noise vector. z The concatenated fingerprint is used as input to the generator, and after passing through the MLP and FC layers, it outputs the enhanced fingerprint corresponding to the reference point. ;

[0030] Step S122: Obtain the discriminator input variables, wherein the discriminator input includes: data from the offline fingerprint database. Real fingerprints and unique hot codes sampled in the middle C The concatenated vector and the enhanced fingerprint generated by the generator based on random noise With unique hot code C A vector formed by concatenation;

[0031] Step S123: After inputting the discriminator input variables into the MLP and FC layers, the discriminator outputs one-dimensional real-valued scores for both real and enhanced fingerprints.

[0032] Step S124: Calculate the loss function of the discriminator. :

[0033]

[0034] Loss function of generator :

[0035]

[0036] in, z It follows a standard normal distribution. A random noise vector, This represents the distribution of real fingerprints. To be from the distribution Real samples from the middle;

[0037] Step S125: Iteratively update the network parameters using the backpropagation algorithm (BP) to optimize the loss function. and Minimize the loss function until it reaches a preset condition, then end CWGAN training and save the CWGAN model parameters.

[0038] Preferably, step S12 further includes:

[0039] Step S126: Concatenate the one-hot code and random noise vector of each reference point and input them into CWGAN to obtain the enhanced fingerprint corresponding to the reference point;

[0040] Step S127: Use the coordinates of the reference points as the coordinates of the corresponding enhanced fingerprints, merge the enhanced fingerprints of all reference points, and form the second offline fingerprint database.

[0041] Preferably, step S13 includes:

[0042] Based on the collection time and generation order, the first offline fingerprint database and the second offline fingerprint database The fingerprint segmentation length of each reference point is s Several sequences, and the first offline fingerprint database and the second offline fingerprint database Merge to obtain a merged offline fingerprint database .

[0043] Preferably, step S14 includes:

[0044] Step S141: Sample the merged offline fingerprint database to obtain the input sequence. The input sequence is fed into a CNN model and an LSTM model, respectively. The CNN model outputs a one-dimensional vector, and the LSTM model updates the cell state. and hidden state ;

[0045] Step S142: Concatenate the output of the last time step of the LSTM model with the one-dimensional vector, and output two-dimensional coordinates based on the multilayer perceptron. ;

[0046] Step S143: Calculate the gradient based on the backpropagation time-based BPTT method for the loss function. Solve for the gradients of the network parameters;

[0047] Step S144: Update the model weight matrix and bias parameters based on the Adam optimization algorithm.

[0048] Preferably, step S14 further includes:

[0049] The loss function The root mean square error of the positioning coordinates is expressed as:

[0050] ,

[0051] in, For batch size, For the first i Predicted coordinates of points , For the first i The true coordinates of each point ;

[0052] When the loss function If the value does not decrease for more than a preset number of consecutive times, it is determined that the parameters of the CNN-LSTM model meet the preset conditions, and the training of the CNN-LSTM model ends.

[0053] This invention can bring at least one of the following beneficial effects:

[0054] This invention proposes a fingerprint database localization method based on CWGAN and neural networks, applied to 5G indoor scenarios. It overcomes signal fluctuations by filtering high-frequency components from samples in the fingerprint database. Then, it processes the original fingerprint database using a CWGAN model, automatically expanding the number of fingerprints to meet requirements and generating enhanced fingerprints with a feature distribution similar to the original fingerprints, thus obtaining an enhanced fingerprint database. After merging the filtered and enhanced fingerprint databases, the merged database is input into a first localization model that has completed training. The first localization model learns the implicit relationships between fingerprint features and their temporal sequence and performs localization. By integrating fingerprint information collected from multiple steps, the method effectively improves the fluctuation of localization results, reduces localization errors, and significantly improves localization accuracy. Simultaneously, it reduces the overhead and cost of fingerprint data collection, thereby improving localization efficiency. Attached Figure Description

[0055] The preferred embodiments will now be described in a clear and easy-to-understand manner, with reference to the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods.

[0056] Figure 1 A schematic diagram illustrating the steps of a fingerprint database localization method based on CWGAN and neural networks, provided in an embodiment of the present invention;

[0057] Figure 2 A top-down view of an indoor positioning scene;

[0058] Figure 3 This is a schematic diagram of the CWGAN model structure in an embodiment of the present invention;

[0059] Figure 4 This is a schematic diagram of the CNN-LSTM model structure in an embodiment of the present invention;

[0060] Figure 5 This diagram illustrates a comparison of the cumulative distribution function of the positioning error between the fingerprint database positioning method based on CWGAN and neural networks provided by this invention and other methods. Detailed Implementation

[0061] The various aspects of the present invention will be further described in detail below.

[0062] Unless otherwise defined or stated, all technical and scientific terms used herein have the same meaning as are familiar to those skilled in the art. Furthermore, any methods and materials similar to or equivalent to those described herein may be used in this invention.

[0063] The following explains the terminology.

[0064] Unless otherwise expressly specified and limited, the term "or" as used in this invention includes the relationship of "and". "And" is equivalent to the Boolean logic operator "AND", and "or" is equivalent to the Boolean logic operator "OR", with "AND" being a subset of "OR".

[0065] It is understood that although the terms "first," "second," etc., may be used herein to describe different elements, these elements should not be limited by these terms. These terms are merely used to distinguish one element from another. Therefore, a first element may be referred to as a second element without departing from the teachings of the inventive concept.

[0066] In this invention, the terms "containing," "comprising," or "including" indicate that various ingredients may be used together in the mixtures or compositions of this invention. Therefore, the term "consistent primarily of..." is included in the terms "containing," "comprising," or "including."

[0067] Unless otherwise expressly specified and limited, the terms "connected," "linked," and "connected" in this invention should be interpreted broadly. For example, they can refer to a fixed connection, a connection through an intermediary, or a connection within two elements or an interaction between two elements. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0068] For example, if an element (or component) is referred to as being on, coupled to, or connected to another element, then the element may be directly formed on, coupled to, or connected to the other element, or there may be one or more intermediate elements between them. Conversely, if the expressions "directly on," "directly coupled to," and "directly connected to" are used herein, then it indicates that there is no intermediate element. Other terms used to describe relationships between elements should be interpreted similarly, such as "between" and "directly between," "attached" and "directly attached," "adjacent" and "directly adjacent," etc.

[0069] Additionally, it should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings. The terms "inner" and "outer" refer to directions toward or away from the geometric center of a particular component, respectively. It is understood that these terms are used here to describe the relationship of one element, layer, or region relative to another element, layer, or region, as shown in the accompanying drawings. These terms should also encompass other orientations of the device in addition to those described in the accompanying drawings.

[0070] Other aspects of the invention will be apparent to those skilled in the art from the disclosure herein.

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0072] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show components related to this application and are not drawn according to the actual number, shape, and size of the components in the actual implementation. In the actual implementation, the form, quantity, and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex. For example, the thickness of the elements in the drawings may be exaggerated for clarity.

[0073] Example

[0074] To address the issues of low accuracy, large fluctuations, and low efficiency in existing 5G indoor positioning technologies, such as... Figure 1 As shown, this invention provides a deep learning fingerprint database localization method based on CWGAN, applied to 5G indoor scenarios, including:

[0075] Step S11: Obtain and filter the offline fingerprint database to obtain the filtered offline fingerprint database. Normalize the filtered offline fingerprint database to obtain the first offline fingerprint database. The fingerprints constituting the offline fingerprint database are the SS-RSRP values ​​of the synchronization signal reference received power of the 5G base station primary and neighboring cells collected by the device in the positioning area.

[0076] Step S12: Process the first offline fingerprint database based on CWGAN. The CWGAN learns the feature distribution of the fingerprints in the first offline fingerprint database and generates a new fingerprint that is similar to the feature distribution of the fingerprints in the first offline fingerprint database. The generated new fingerprint constitutes the second offline fingerprint database.

[0077] Step S13: Merge the first offline fingerprint database and the second offline fingerprint database to obtain a merged offline fingerprint database;

[0078] Step S14: Obtain the CNN-LSTM model. Input the merged offline fingerprint database into the CNN-LSTM model. When the parameters of the CNN-LSTM model meet the preset conditions, complete the training of the CNN-LSTM model and obtain the first localization model.

[0079] Step S2: Collect fingerprints at the test points and process the data according to step S1;

[0080] Step S3: Input the processed fingerprint of the test point into the first positioning model, and the first positioning model outputs the predicted coordinate value of the test point.

[0081] In one specific embodiment of the present invention, obtaining the offline fingerprint database in step S11 includes:

[0082] Step S111: Divide the positioning area into N The number of grids and the number of 5G base stations within the positioning area are: M Each grid intersection is designated as a reference point. The fingerprint of each reference point is obtained, and the fingerprint feature vectors of all reference points constitute an offline fingerprint database. , ;

[0083] in, It is the first i A reference point at a certain time t The fingerprint formed by the SS-RSRPs of all 5G base stations received internally, 1≤ i≤N .

[0084] In this embodiment, the indoor scene is as follows: Figure 2 As shown, the room size is 14m × 9m, and six 5G base stations (AP1, AP2, AP3, AP4, AP5, and AP6) are deployed. The room is divided into a grid with 1m intervals, and 5G devices are fixed at the intersections of each grid. The RSRP value within the collection time T is used as the fingerprint of each grid intersection. In step S1, filtering the data can effectively overcome the fluctuation of the received signal and improve the positioning accuracy. In step S2, CWGAN means using a conditional Wasserstein generative adversarial network to learn the fingerprint feature distribution of different grids, generating corresponding enhanced fingerprints on the basis of the original limited dataset, thereby expanding the dataset and reducing the manpower consumption and collection cost of data collection.

[0085] In one specific embodiment of the present invention, filtering the offline fingerprint database in step S11 includes:

[0086] Step S112, the i The feature vector of a fingerprint from a single sample at each reference point is represented as follows: ,delete Vectors containing null values ​​are selected, and those with SS-RSRP values ​​higher than a threshold are filtered out. The sample is used as the fingerprint of this reference point. ;

[0087] Step S113: Obtain the time tThe extreme points of the received fingerprint are identified, and the upper envelope is obtained by curve interpolation based on these extreme points. and lower envelope and the upper envelope and the lower envelope Calculate the average to obtain the average envelope. And calculate the remaining components. :

[0088]

[0089] Step S114, Repeat step S113 as a new signal until the difference between the number of extreme points and zero-crossing points is no greater than 1, and the average value of the envelope of the maximum and minimum values ​​at any point is 0. At this point, the remaining components... Components of the first-order eigenmode function ;

[0090] Step S115: Calculate the first-order residual: ;

[0091] make Alternative Repeat steps S113 to S114 until the residual is monotonic or no longer oscillates.

[0092] Step S116, Discard the previous steps p The first-order intrinsic mode function component is obtained by summing all the remaining intrinsic mode function components to obtain the filtered fingerprint of the reference point. n is the total number of intrinsic mode function components;

[0093] Step S117: Merge all reference points Obtain the filtered offline fingerprint database .

[0094] In one specific embodiment of the present invention, step S11 further includes:

[0095] Step S118: Filter the offline fingerprint database Normalize each column, the first... k The normalized SS-RSRP value is expressed as follows: , Indicates the first k The minimum value of SS-RSRP. Indicates the first k The maximum value of SS-RSRP in each column is used to merge the results of each column to form the normalized first offline fingerprint database. , normed This indicates normalization processing;

[0096] Step S119: Use one-hot encoding to process the numbering of each reference point. One-hot codes for all corresponding reference points: ,in, For the first i One-hot codes corresponding to each sample The total number of samples.

[0097] In one specific embodiment, p =2, which means filtering out the first and second order modal components.

[0098] This network adds prior information of one-hot codes for reference points to the traditional CWGAN network, enabling the network to generate enhanced fingerprints for specified reference points without training a separate model for each reference point, significantly reducing model training time and enhancing model universality.

[0099] In one specific embodiment of the present invention, such as Figure 3 As shown, CWGAN includes a generator and a discriminator, which are respectively composed of a multilayer perceptron (MLP) and a fully connected (FC) layer. Step S12 includes:

[0100] Step S121: Take the one-hot code C of all grids as a priori conditions, and combine it with the random noise vector. z The concatenation is used as input to the generator, and after passing through the MLP and FC layers, the enhanced fingerprint corresponding to the grid is output. ;

[0101] Step S122: Obtain the discriminator input variables, wherein the discriminator input includes: data from the offline fingerprint database. The vector formed by concatenating the real fingerprint sampled in the middle with the grid one-hot code C, and the enhanced fingerprint generated by the generator based on random noise. A vector formed by concatenating the grid one-hot code C;

[0102] Step S123: After inputting the discriminator input variables into the MLP and FC layers, the discriminator outputs one-dimensional real-valued scores for both real and enhanced fingerprints. or ;

[0103] Step S124: Calculate the loss function of the discriminator. :

[0104] ,

[0105] Loss function of generator :

[0106] ,

[0107] in,z Follows a standard normal distribution , This represents the distribution of real fingerprints. To be from the distribution Real samples from the middle;

[0108] Step S125: Iteratively update the network parameters using the backpropagation algorithm (BP) to optimize the loss function. and Minimize the loss function until the third preset condition is met, then end CWGAN training and save the CWGAN model parameters.

[0109] In one specific embodiment of the present invention, step S12 further includes:

[0110] Step S126: Concatenate the one-hot encoding and random noise vector of each grid and input them into CWGAN to obtain the enhanced fingerprint of each grid.

[0111] Step S127: Use the grid point coordinates as the coordinates of the corresponding grid's enhanced fingerprint, merge the enhanced fingerprints of all grids to form the second fingerprint database. .

[0112] In a specific application example, the generator network consists of 5 units connected in series, each unit consisting of two hidden layers with 30 to 20 neurons, with the activation function between the hidden layers being LeakyReLU and the learning rate being 0.001; the discriminator MLP consists of two hidden layers with 50 neurons each, with the activation function between the hidden layers being LeakyReLU and the learning rate being 0.0005.

[0113] In a preferred embodiment of the present invention, step S13 includes:

[0114] Based on the collection time and generation order, the first fingerprint database and the second fingerprint database The fingerprint segmentation length of each grid is s Several sequences, and the first fingerprint database and the second fingerprint database Merge to obtain a merged fingerprint database .

[0115] Furthermore, the fingerprint databases will be merged at an 8:2 ratio. The dataset is divided into training and validation sets to train the generator and discriminator. This expands the dataset without requiring repeated manual data collection, saving manpower and time costs and improving the model's prediction accuracy.

[0116] In one specific embodiment of the present invention, the CNN-LSTM model consists of a one-dimensional convolutional layer (CNN1d) and a long short-term memory network (LSTM). The CNN1d layer consists of multiple convolutional kernels, and the LSTM consists of multiple LSTM units. Each unit includes a cell state, a forget gate, an input gate, and an output gate. Step S14 includes:

[0117] Step S141: Sample the merged offline fingerprint database to obtain the input sequence. The input sequence is fed into a CNN model and an LSTM model, respectively. The CNN model outputs a one-dimensional vector, and the LSTM model updates the cell state. and hidden state ;

[0118] Step S142: Concatenate the output of the last time step of the LSTM model with the one-dimensional vector, and output two-dimensional coordinates based on the multilayer perceptron. ;

[0119] Step S143: Calculate the gradient based on the backpropagation time-based BPTT method for the loss function. Solve for the gradients of the network parameters;

[0120] Step S144: Update the model weight matrix and bias parameters based on the Adam optimization algorithm.

[0121] Specifically, the one-dimensional vector output by the CNN model includes:

[0122] The convolution kernel performs a one-dimensional convolution operation along the feature dimension, and the output of the convolutional layer is represented as:

[0123]

[0124] in, It is the nth element of the output sequence, where x is the input fingerprint sequence. Let K be the k-th element of the convolution kernel, where K is the kernel length. The output of the convolutional layer is further processed by a max pooling layer to extract local features of the receptive field, and then flattened by a flattening layer to obtain a one-dimensional vector.

[0125] Specifically, LSTM updates cell state and hidden state This includes updating the cell state using forget gates, input gates, and output gates at each time step. and hidden state .

[0126] In one specific embodiment of the present invention, step S14 further includes:

[0127] The loss function The root mean square error of the positioning coordinates is expressed as:

[0128] ,

[0129] in, For batch size, For the first i Predicted coordinates of points , The true coordinates of the i-th point ;

[0130] When the loss function When the value is not higher than the preset threshold, the training of the CNN-LSTM model ends.

[0131] Specifically, in a practical application example, the convolutional layers are set to 2 layers, the fingerprint feature dimension is 6, the number of channels in the convolutional layers are 10 and 32 respectively, the sequence length s=10, and the batch size for data batching is... =128, learning rate =0.001, the number of neurons in the MLP hidden layer is (128,256,128).

[0132] Figure 5 The diagram shows a comparison of the cumulative distribution function of the positioning error in a 5G indoor scenario using the fingerprint database positioning method based on CWGAN and neural networks provided by this invention (yellow), with the cumulative distribution function using only Multilayer Perceptron (MLP) (red), only CNN-LSTM model (blue), and only Empirical Mode Decomposition (EMD) and CNN-LSTM model (green). The vertical axis represents the cumulative distribution of errors, and the horizontal axis represents the positioning error, with the unit being meters.

[0133] In summary, the present invention achieves the following effects:

[0134] This invention proposes a fingerprint database localization method based on CWGAN and neural networks, applied to 5G indoor scenarios. It overcomes signal fluctuations by filtering high-frequency components from samples in the fingerprint database. Then, it processes the original fingerprint database using a CWGAN model, automatically expanding the number of fingerprints to meet requirements and generating enhanced fingerprints with a feature distribution similar to the original fingerprints, thus obtaining an enhanced fingerprint database. The method merges a first fingerprint database with a second fingerprint database, inputting the merged database into a trained first localization model. The first localization model learns the implicit relationships between fingerprint features and their temporal sequence for localization. By integrating fingerprint information collected from multiple steps, the method effectively improves the fluctuation of localization results, reduces localization errors, and significantly enhances localization accuracy. Simultaneously, it reduces the overhead and cost of fingerprint data collection, thereby improving localization efficiency.

[0135] Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspect set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or practice the method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0136] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0137] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0138] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing description of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A deep learning fingerprint database localization method based on CWGAN, applied to 5G indoor scenarios, characterized in that, include: Step S1: Train the first localization model, including: Step S11: Divide the positioning area into N The number of grids and the number of 5G base stations within the positioning area are: M Each grid intersection is set as a reference point. The fingerprints of each reference point are used to construct an offline fingerprint database. The offline fingerprint database is then filtered to obtain the filtered offline fingerprint database. The filtered offline fingerprint database is normalized to obtain the first offline fingerprint database; the fingerprints constituting the offline fingerprint database are the SS-RSRP values ​​of the synchronization signal reference received power of the primary and neighboring cells of the 5G base station collected by the device in the positioning area. Step S12: Process the first offline fingerprint database based on CWGAN. The CWGAN learns the feature distribution of fingerprints in the first offline fingerprint database and generates new fingerprints similar to the feature distribution of fingerprints in the first offline fingerprint database. The generated new fingerprints constitute the second offline fingerprint database. CWGAN includes a generator and a discriminator, which are respectively composed of a multilayer perceptron (MLP) and a fully connected (FC) layer. Step S12 includes: Step S121: Obtain the unique hot codes for all reference points. C As a priori condition, with random noise vector z The concatenated fingerprint is used as input to the generator, and after passing through the MLP and FC layers, it outputs the enhanced fingerprint corresponding to the reference point. ; Step S122: Obtain the discriminator input variables, wherein the discriminator input includes: data from the offline fingerprint database. Real fingerprints and unique hot codes sampled in the middle C The concatenated vector and the enhanced fingerprint generated by the generator based on random noise With unique hot code C A vector formed by concatenation; Step S123: After inputting the discriminator input variables into the MLP and FC layers, the discriminator outputs one-dimensional real-valued scores for both real and enhanced fingerprints. or ; Step S124: Calculate the loss function of the discriminator. : ， Loss function of generator : ， in, z It follows a standard normal distribution. A random noise vector, This represents the distribution of real fingerprints. To be from the distribution Real samples from the middle; Step S125: Iteratively update the network parameters using the backpropagation algorithm (BP) to optimize the loss function. and Minimize the loss function until it reaches a preset condition, then terminate CWGAN training and save the CWGAN model parameters. Step S126: Concatenate the one-hot code and random noise vector of each reference point and input them into CWGAN to obtain the enhanced fingerprint corresponding to the reference point; Step S127: Use the coordinates of the reference points as the coordinates of the corresponding enhanced fingerprints, merge the enhanced fingerprints of all reference points, and form the second offline fingerprint database. ; Step S13: Merge the first offline fingerprint database and the second offline fingerprint database to obtain a merged offline fingerprint database; Step S13 includes: Based on the collection time and generation order, the first offline fingerprint database and the second offline fingerprint database The fingerprint segmentation length of each reference point is s Several sequences, and the first offline fingerprint database and the second offline fingerprint database Merge to obtain a merged offline fingerprint database ; Step S14: Obtain the CNN-LSTM model. Input the merged offline fingerprint database into the CNN-LSTM model. When the parameters of the CNN-LSTM model meet the preset conditions, complete the training of the CNN-LSTM model to obtain the first localization model. Step S14 further includes: Step S141: Sample the merged offline fingerprint database to obtain the input sequence. The input sequence is fed into a CNN model and an LSTM model, respectively. The CNN model outputs a one-dimensional vector, and the LSTM model updates the cell state. and hidden state ; The CNN model outputs a one-dimensional vector, specifically by performing a one-dimensional convolution operation along the feature dimension using a convolution kernel. Step S2: Collect fingerprints at the test points and process the data according to step S1; Step S3: Input the processed fingerprint of the test point into the first positioning model, and the first positioning model outputs the predicted coordinate value of the test point.

2. The deep learning fingerprint database localization method based on CWGAN as described in claim 1, characterized in that, Step S11, obtaining the offline fingerprint database, includes: Step S111: The fingerprint feature vectors of all reference points constitute an offline fingerprint database. , ; in, It is the first i A reference point at a certain time t The fingerprint formed by the SS-RSRPs of all 5G base stations received internally, 1≤ i≤N .

3. The CWGAN-based deep learning fingerprint database localization method according to claim 2, characterized in that, The filtering of the offline fingerprint database in step S11 includes: Step S112, the i The feature vector of a fingerprint from a single sample at each reference point is represented as follows: ,delete Vectors containing null values ​​are selected, and those with SS-RSRP values ​​higher than a threshold are filtered out. The sample is used as the fingerprint of this reference point. ; Step S113: Obtain the time t The extreme points of the received fingerprint are identified, and the upper envelope is obtained by curve interpolation based on these extreme points. and lower envelope and the upper envelope and the lower envelope Calculate the average to obtain the average envelope. And calculate the remaining components. : ； Step S114, Repeat step S113 as a new signal until the difference between the number of extreme points and zero-crossing points is no greater than 1, and the average value of the envelope of the maximum and minimum values ​​at any point is 0. At this point, the remaining components... The first-order eigenmode function components ; Step S115: Calculate the first-order residual: ; make Alternative Repeat steps S113 to S114 until the residual is monotonic or no longer oscillates. Step S116, Discard the previous steps p The first-order intrinsic mode function component is obtained by summing all the remaining intrinsic mode function components to obtain the filtered fingerprint of the reference point. n is the total number of intrinsic mode function components; Step S117: Merge all reference points Obtain the filtered offline fingerprint database .

4. The CWGAN-based deep learning fingerprint database localization method according to claim 3, characterized in that, Step S11 also includes: Step S118: Filter the offline fingerprint database Normalize each column, the first... k The normalized SS-RSRP value is expressed as follows: , Indicates the first k The minimum value of SS-RSRP. Indicates the first k The maximum value of SS-RSRP in each column is used to merge the results of each column to form the normalized first offline fingerprint database. , normed This indicates normalization processing; Step S119: Use one-hot encoding to process the numbering of each reference point. One-hot codes for all corresponding reference points: ,in, For the first i One-hot codes corresponding to each sample The total number of samples.

5. The CWGAN-based deep learning fingerprint database localization method according to claim 4, characterized in that, Step S14 further includes: Step S142: Concatenate the output of the last time step of the LSTM model with the one-dimensional vector, and output two-dimensional coordinates based on the multilayer perceptron. ; Step S143: Calculate the gradient based on the backpropagation time-based BPTT method for the loss function. Solve for the gradients of the network parameters; Step S144: Update the model weight matrix and bias parameters based on the Adam optimization algorithm.

6. The CWGAN-based deep learning fingerprint database localization method according to claim 5, characterized in that, Step S14 also includes: The loss function The root mean square error of the positioning coordinates is expressed as: ， in, For batch size, For the first i Predicted coordinates of points , For the first i The true coordinates of each point ; When the loss function If the value does not decrease for more than a preset number of consecutive times, it is determined that the parameters of the CNN-LSTM model meet the preset conditions, and the training of the CNN-LSTM model ends.

Images