Fingerprint positioning method and device based on deep convolutional neural network, and medium

Abstract

The invention discloses a fingerprint positioning method and device based on a deep convolutional neural network, and a medium. The method comprises an offline stage and an online stage. The offline stage comprises the following steps: dividing a target positioning region into a plurality of sub-regions, and allocating different numbers to each region; selecting a plurality of reference points in each sub-region, measuring and collecting fingerprint matrixes of all the reference points, and constructing a data set; building and training a deep convolutional neural network model; the online stage comprises the following steps: measuring test points in a target positioning area to obtain a fingerprint matrix; inputting the fingerprint matrix of the test points into a deep convolutional neural network model to obtain sub-region numbers; and calculating position fingerprints of all reference points in the corresponding sub-regions, calculating similarity distances between the position fingerprints of all the reference points and the sub-region serial number position fingerprints, selecting a plurality of reference points according to the similarity distances, and calculating according to a WKNN algorithm to obtain specific position coordinates corresponding to the fingerprints of the test points so as to realize fingerprint positioning.

Description

Technical Field

[0001] The present invention relates to the technical field of communication and information systems, and in particular to a fingerprint positioning method, device and medium based on a deep convolutional neural network. Background Art

[0002] Fingerprint positioning technology uses the influence of multipath effect on signals at different locations to collect a sufficient number of "signal fingerprints" in a specific area as a location reference to achieve the positioning of mobile users. Compared with traditional wireless network positioning technology, the main advantage of fingerprint positioning is that its positioning accuracy is not affected by multipath effect, and it can make full use of multipath information to enhance the characteristic parameters of "signal fingerprint".

[0003] Due to the multipath effect, wireless signals will experience reflection, diffraction and other phenomena when transmitted from mobile terminals to base stations, resulting in the signal received by the base station being the superposition of multiple path signals. Therefore, signals from different locations pass through different propagation paths, and the signal parameters received by the base station will be different. This is the main principle of implementing fingerprint positioning technology.

[0004] The workflow of fingerprint positioning is divided into two steps: offline and online. In the offline stage, the target area to be positioned is first divided into grids at appropriate intervals, and reference nodes are selected. Then, each reference grid point is traversed to extract the signal fingerprint information (such as CSI) that can represent the location, and these location fingerprint information and the location coordinates of the reference point are stored in the fingerprint library in a one-to-one correspondence. The task of the online stage is to estimate the location coordinates of the terminal to be positioned. The location fingerprint of the user to be positioned is measured and obtained in the same way as the reference point location fingerprint, and then the "distance" between the point to be positioned and the grid reference point is calculated based on the selected fingerprint similarity criterion, and finally the matching algorithm is used to estimate its true coordinates. Summary of the invention

[0005] Technical purpose: In view of the defects in the prior art, the present invention discloses a fingerprint positioning method, device and medium based on deep convolutional neural network, which can not only significantly reduce the positioning error, but also greatly shorten the positioning time in the online stage, meeting the real-time requirements of the user to be located.

[0006] Technical solution: In order to achieve the above technical objectives, the present invention adopts the following technical solution.

[0007] A fingerprint positioning method based on a deep convolutional neural network, the method comprising an offline stage and an online stage; the offline stage comprises the following steps: S1.1, dividing the target positioning area into several sub-areas, and assigning different numbers to each area; S1.2. Select several reference points in each sub-area, measure and collect the ADCFA matrices of all reference points as fingerprint matrices; construct a data set with the fingerprint matrices, position coordinates and sub-area numbers of all reference points; S1.3. Build a deep convolutional neural network model and train it using the data set to obtain a trained deep convolutional neural network model; The online phase includes the following steps: S2.1, measure the test points in the target positioning area to obtain a fingerprint matrix; S2.2, input the fingerprint matrix of the test point into the deep convolutional neural network model trained in the offline stage, and output the predicted sub-region number; S2.3. Calculate the position fingerprints of all reference points in the sub-region corresponding to the sub-region number, calculate the similarity distance between the position fingerprints of all reference points and the position fingerprint of the sub-region number, select several reference points according to the similarity distance, and calculate the specific position coordinates corresponding to the fingerprint of the test point according to the WKNN algorithm to achieve fingerprint positioning.

[0008] An electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program so that the device performs the above-mentioned fingerprint positioning method based on a deep convolutional neural network.

[0009] A computer storage medium having a computer program stored thereon. When the computer program is executed, a device executing the computer program implements the above-mentioned fingerprint positioning method based on a deep convolutional neural network.

[0010] Beneficial effect: The present invention divides the fingerprint positioning process into an offline stage and an online stage. In the offline stage, when extracting the fingerprint matrix, the unitary matrix is ​​discarded and two new fingerprint matrices are obtained by taking the square root of each element. This not only significantly reduces the computational complexity, but also greatly shortens the positioning time in the online stage, thus meeting the real-time requirements of the user to be positioned. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of the fingerprint positioning method in Example 1; Figure 2 This is a schematic diagram of coordinate points of the regional classification example in Example 2; Figure 3 This is a schematic diagram of ADCFP amplitude distribution during fingerprint matrix extraction in Example 2; Figure 4 Schematic diagram of ADCFA amplitude distribution during fingerprint matrix extraction in Example 2; Figure 5This is a schematic diagram of the visualization of the DCNN network model structure in Example 2. DETAILED DESCRIPTION

[0012] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0013] As attached Figure 1 As shown, the present invention discloses a fingerprint positioning method, device and medium based on a deep convolutional neural network, including two key steps: the construction of a fingerprint matrix and a fingerprint positioning method based on a deep convolutional neural network. The construction process of the fingerprint matrix includes the following steps: first, the base station obtains the CIR and CFR matrices of the wireless terminal signal through channel estimation. The CFR matrix is ​​subjected to DFT transformation to obtain the angle delay domain channel response matrix ADCFR, the ADCFR matrix is ​​subjected to Hadamard inner product operation to obtain the angle delay domain channel energy matrix ADCFP, and the ADCFP matrix elements are obtained by taking the square root to obtain the angle delay domain channel amplitude matrix ADCFA. The fingerprint positioning method based on a deep convolutional neural network includes the following steps: dividing the positioning sub-area and assigning regional labels, extracting the fingerprint information of the grid points to construct a fingerprint library, building a DCNN network model, training the model using the fingerprint information of the grid points and the regional labels, performing regional classification on the fingerprint of the user to be tested in the online stage, and using the WKNN algorithm to perform position estimation to obtain specific position coordinates. The present invention can not only significantly reduce the positioning error, but also greatly shorten the positioning time in the online stage, meeting the real-time requirements of the user to be positioned. The following is a detailed description of the steps of a fingerprint positioning method based on a deep convolutional neural network according to the present invention.

[0014] Example 1 like Figure 1 As shown, a fingerprint positioning method based on a deep convolutional neural network in this embodiment includes the following steps: S1, Offline Phase: S1.1, dividing the target positioning area in the spatial coordinate system into a number of sub-areas with equal distances from the x direction and the y direction, and assigning different numbers to each area in sequence; S1.2. In each sub-area, the grid is evenly divided at fixed intervals and several reference points are selected. The reference points are evenly distributed in the target positioning area and have corresponding position coordinates in the spatial coordinate system. The ADCFA matrix at the positions of all reference points is measured and collected as the fingerprint matrix; the fingerprint matrix, position coordinates and sub-area numbers of all reference points are used to construct a data set.

[0015] The fingerprint matrix is ​​constructed as follows: S1.2.1. There is a base station in the target positioning area. The base station obtains the user through channel estimation. The CFR matrix of the signal sent to the base station is recorded as ;user Locate any point in the target area; S1.2.2. Perform DFT transformation to obtain the angle delay domain channel response matrix, namely the ADCFR matrix, which is recorded as ; The calculation formula includes: , in, for The unitary matrix of is the phase-shift DFT matrix, is the number of antennas of the base station, The rows and columns are , The set of complex matrices of , is the number of subcarriers that send signals to the user; the calculation formulas for the components in the phase-shifted DFT matrix are: , in, The rows and columns are The square array, For square The element corresponding to the mth row and nth column in ; S1.2.3. ADCFR Matrix Calculate the Hadamard inner product with its own conjugate matrix to obtain the angle delay domain channel energy matrix, that is, the ADCFP matrix, recorded as ; The calculation formula includes: , in, is the mean operator, For the ADCFR matrix The conjugate matrix of The rows and columns are The set of real matrices of , for The element corresponding to the mth row and nth column in ; S1.2.4. ADCFP Matrix By taking the square root of all its elements, we can obtain the angle delay domain channel amplitude matrix, namely the ADCFA matrix, which is recorded as ; The calculation formula includes: , in, For the ADCFP matrix The conjugate matrix of The rows and columns are The set of real matrices of , for The element corresponding to the mth row and nth column in .

[0016] The fingerprint matrix obtained by this method not only maintains the relative size relationship between the elements, but also effectively reduces the value of each element, thereby greatly reducing the calculation and storage overhead of the fingerprint matrix and improving the processing speed of the online positioning stage.

[0017] According to the grid calculation step size, the x- and y-coordinates of the reference point are converted to the corresponding equidistant segments, and the classes are assigned to the data set based on the grid, such as the grid (1,1) is assigned to the class class = 1, and the grid (m,n) is assigned to the class class = m×n. The data consisting of (ADCFA, class) pairs are randomly divided into 90% training data sets and 10% test data sets, where class represents the sub-region number.

[0018] S1.3. Build a deep convolutional neural network model, divide the data set into a training data set and a test data set, use the fingerprint matrix of each reference point as training data, and the number of the sub-region where the reference point is located as the training label to train the deep convolutional neural network model. When the loss function value is reduced to the expected range, save the parameter weights and various biases of each node of the deep convolutional neural network model. The deep convolutional neural network model is used for online prediction.

[0019] The deep convolutional neural network model includes: The input layer is used to normalize the input data to obtain normalized data with a dimension of , is the number of antennas of the base station; the input data is the fingerprint matrix; Five groups of convolutional layers and pooling layers with the same structure, where: Convolution layer, used to perform convolution operation on the normalized data and output the convolution data; there are 5 convolution layers in total, the first convolution layer has 8 convolution kernels, the second convolution layer has 16 convolution kernels, the third convolution layer has 32 convolution kernels, the fourth convolution layer has 64 convolution kernels, and the fifth convolution layer has 128 convolution kernels; The pooling layer is used to perform maximum pooling on the convolutional data to obtain the pooled data; The fully connected layer is used to perform maximum pooling on the pooled data to obtain the predicted sub-region number; S2, online stage: S2.1. Randomly select test points in the target positioning area and measure the fingerprint matrix of these test points; S2.2, input the fingerprint matrix measured in S2.1 into the deep convolutional neural network model trained in the offline stage, output the probability of the fingerprint matrix being located in different sub-regions through the trained deep convolutional neural network model, compare the probability values ​​and use the sub-region number with the largest probability value as the output result of the trained deep convolutional neural network model; S2.3. Calculate the similarity distances between all reference points and the test point in the sub-region corresponding to the sub-region number, select K reference points with the closest similarity distances from these reference points, K=3 in this embodiment; and estimate the specific position coordinates corresponding to the fingerprint of the test point according to the WKNN algorithm to achieve fingerprint positioning.

[0020] The similarity distance is calculated using the cosine distance, and the calculation formula is: , in, for arrive The vector of the point, for arrive The vector of the point, for The F-norm of for The F-norm of Represents vectorization; For the base site, , They are test point and reference point respectively; The calculation formula of the WKNN algorithm includes: , in, is the position coordinate obtained using the weighted K nearest neighbor algorithm, is the position coordinate of the kth reference point, is the horizontal and vertical coordinates; is the weight corresponding to the kth reference point and satisfies , is the number of reference points, is the fingerprint distance between the fingerprint matrix at the kth reference point and the fingerprint matrix at the target position to be located, is a positive number. Since these distances may be very small, in order to avoid the situation where the denominator is close to 0, a small positive number is added here ; The value is selected according to the actual situation.

[0021] By introducing the weighted idea, points that are closer to the fingerprint similarity of the target terminal to be located are assigned larger weights, and therefore have a more significant impact on the location estimation result; conversely, points that are farther away from the fingerprint similarity of the target terminal to be located are assigned smaller weights, and therefore have less impact on the location estimation result.

[0022] Example 2 A fingerprint positioning method based on a deep convolutional neural network in this embodiment includes the following steps: like Figure 1 As shown, it includes two major steps: offline stage and online stage: Step 1: Offline phase: 1) Divide into sub-areas and assign labels: (a) determining the overall shape and size of the target location area, dividing the target location area into multiple sub-areas according to preset requirements, and assigning numbers according to the coordinate sequence; (b) Taking the target positioning area as a rectangular area as an example, the length of the positioning area is 5.5 meters, the width is 1.2 meters, and the height is 0.6 meters. In this paper, BS2 is selected as the base station. BS2 is located on the left side of the positioning area and is installed 2 meters above the ground; (c) Five classification types were used to divide the positioning area into 1 5, 6 10, 8 20, 12 40, and 18 55 sub-areas, and five groups of comparative tests were conducted. The first group was divided into 1 on the horizontal axis and 5 on the vertical axis, and so on; the results of these five types were compared.

[0023] 2) Select grid reference points and measure fingerprint matrix information: (a) In Figure 2 After the target positioning area is divided into sub-areas, each sub-area is evenly divided into grids with an interval of 0.01 meters. In this embodiment, the target positioning area is divided into 551 rows, each row has 121 reference points. The total number of grid reference points selected in this way is 66671, which are distributed in different sub-areas.

[0024] (b) At the grid reference point, the user terminal is responsible for sending signals, while the base station uses channel estimation technology to capture and obtain the CFR (channel frequency response) matrix corresponding to the signal sent by the user to the base station.

[0025] (c) For the obtained CFR matrix, we perform DFT (discrete Fourier transform) according to the formula in Example 1 to calculate the angle delay channel frequency response matrix (ADCFR). Next, we calculate the Hadamard inner product of this ADCFR matrix and its own conjugate matrix to obtain the angle delay channel frequency power matrix (ADCFP).

[0026] (d) To further process the ADCFP matrix, we first calculate the square root of all its elements to obtain the angle delay channel frequency amplitude matrix (ADCFA). This ADCFA matrix constitutes the ADCFA fingerprint at a specific grid reference point, also known as the fingerprint matrix. In order to obtain the location fingerprint information of the entire area, we will continue to traverse all grid reference points and repeat the above process to obtain the ADCFA fingerprint information at all grid reference points.

[0027] (e) Taking the case where the number of antennas is 128 and the number of OFDM cyclic prefixes is 144 as an example, the amplitude distribution of the fingerprint matrix can be described as follows: Figure 3 and Figure 4 The three-dimensional bar graph shown shows the amplitude distribution of the fingerprint matrix. The figure is presented in the form of a bar graph with the columns of the fingerprint matrix (representing the arrival time (ToA) components of different paths) as the X-axis, the rows of the matrix (representing the arrival angle (AoA) components of different paths) as the Y-axis, and the amplitude or energy of the fingerprint matrix as the Z-axis. Since the element values ​​of ADCFA (the processed amplitude-related fingerprint matrix) are obtained by taking the square root of the element values ​​in ADCFP (the original amplitude-related fingerprint matrix), the relative magnitude relationship between the elements remains the same despite the difference in the two in terms of their representation. Specifically, the amplitude near the 50th AoA component in the angle domain is the highest, while the amplitudes in other areas on both sides are relatively low. However, there are some special values ​​in the ADCFP matrix, whose values ​​are huge and significantly different from the values ​​of most elements in the matrix by more than 3 orders of magnitude. The existence of these outliers may bring additional computational burden and complexity to the subsequent calculation process. In contrast, since the element values ​​in the ADCFA matrix have been processed by taking square roots, the differences between the element values ​​are effectively reduced, making the subsequent calculation process more efficient and convenient.

[0028] 3) The ADCFA fingerprint matrix at all grid reference points, the corresponding position coordinates, and the sub-region numbers to which they belong are completely stored in the database to prepare for subsequent model training and position estimation. In the model training stage, these data will serve as important input features to help the algorithm learn and build an accurate positioning model.

[0029] 4) Combined with the deep convolutional neural network DCNN model, the fingerprint matrix at the grid reference point position is used as the training data, and the number of the sub-region where the reference point is located is used as the training label to train the network model until the loss function value is reduced to an acceptable range, and then the network model node parameter weights and various biases are saved for online prediction. Figure 5 As shown, the basic structure of the DCNN model is as follows: (a) Input layer: The input data of the network model is the ADCFA matrix at each grid reference point. Before the data enters the convolution layer, it will first be preprocessed by normalization to ensure the stability and efficiency of model training. Taking the fingerprint data obtained when the number of antennas is 32 and the number of OFDM cyclic prefixes is 144 as an example, the dimension of the fingerprint matrix at this time corresponds to the number of antennas, that is, it has 32 feature dimensions. These preprocessed fingerprint matrices will be used as input data of the network model for feature extraction and learning in subsequent convolution layers.

[0030] (b) Convolutional layer (Conv): A total of five convolutional layers are added to the DCNN network model. The input of the first convolutional layer is the preprocessed fingerprint matrix. There are 8 convolutional kernels of size 16×16×1 in this layer. The sizes of the convolutional kernels from the second to the fifth layers are 8×8×4 (16), 7×7×8 (32), 5×5×16 (64), and 3×3×32 (128), respectively. Each layer uses the ReLU activation function for feature extraction, generates a feature map through convolution operation, and passes it to the next layer; (c) Pooling layer: In the DCNN network model, max pooling is used to further compress the feature map output by the convolution layer. The pooling window size is 2×2, and the regions are divided in a non-overlapping manner. For each divided region, the maximum value is taken as the representative of the region, that is, a downsampling operation is performed. Taking the data matrix dimension of 32×32×8 as an example, after max pooling, the dimension of the data matrix will change to 16×16×8. This operation can reduce the dimension of the feature map and reduce the amount of calculation while retaining the main feature information. After that, the convolution layer and pooling layer operations are repeated four times to further extract and compress the features.

[0031] (d) Fully connected layer: After the convolution and pooling operations, the fully connected layer integrates the previously extracted feature information: (d1) After the DCNN network model undergoes five convolution and pooling operations, the dimension of the data matrix is ​​compressed to 2×2×128. In order to effectively map the features extracted at this level to the sample space, the model will perform a flattening operation, that is, converting this three-dimensional matrix into a one-dimensional vector, so that its dimension becomes 128×1. Subsequently, through a fully connected layer (Dense Layer) containing 128 neurons, the network will integrate these feature information for subsequent classification tasks.

[0032] (d2) Perform Dropout operation on each network layer with a dropout probability of 0.5 to reduce the risk of overfitting and increase the robustness of the model; (d3) At the end of the DCNN network model, a fully connected layer is set, and its length matches the number of sub-regions divided. For the application scenario with 5 region numbers, the length of the fully connected layer is set to 5, which is used to receive the feature information integrated after the previous convolution and pooling operations. In terms of loss function, the cross entropy loss function commonly used in classification models is selected, which can effectively measure the difference between the probability distribution predicted by the model and the true label. In the model prediction stage, for the input fingerprint matrix, the network will calculate the probability size of it belonging to each of the five sub-regions. This process is implemented through the Softmax classifier, which converts the raw score output by the network into a probability distribution and ensures that the sum of the probabilities of all sub-regions is 1. Finally, the model will select the region label with the highest probability as the prediction result and as the final output of the model.

[0033] Step 2: Online stage: 5) According to the aforementioned fingerprint matrix acquisition method, the fingerprint matrix of the mobile user terminal at a random position in the positioning area is measured; 6) Input the measured fingerprint matrix into the DCNN network model trained in the offline stage, calculate the probability of it being located in different sub-regions, and select the sub-region category with the largest probability value as the model output result according to the probability; 7) According to the judgment result of the network model, the similarity distance between the fingerprint matrix of all grid reference points in the sub-area and the fingerprint matrix of the mobile terminal to be located is calculated. The similarity distance is calculated using the cosine distance. The larger the cosine distance, the farther the distance between the two points in the multidimensional space, and the lower the similarity. The smaller the cosine distance, the closer the distance between the two points in the multidimensional space, and the higher the similarity.

[0034] 8) According to the calculation result of the cosine distance, select the K reference points with the closest similarity distance to the target terminal to be located from all the reference grid points in the sub-area, and estimate the specific location coordinates of the user according to the WKNN algorithm. Due to the introduction of the weighted idea, the points with closer similarity distance to the target terminal fingerprint to be located are assigned a larger weight, and the greater the impact on the location estimation result; on the contrary, the points with farther similarity distance to the target terminal fingerprint to be located are assigned a smaller weight, and the smaller the impact on the location estimation result.

[0035] The present invention can reduce the positioning error while reducing the positioning time required in the online stage, meeting the real-time requirements of the user to be positioned. First, the fingerprint matrix is ​​selected. Based on the ADCPM matrix in the angle delay domain, an ADCFA matrix with simpler operation and more significant position discrimination is proposed. It shows a relatively balanced channel distribution on both strong and weak paths, and the positioning accuracy is the highest under different numbers of antennas. The average positioning error is reduced by about 80.43% compared with the ADCPM matrix. Under the same storage overhead, the fingerprint database generation time is about 1 / 10 of the ADCPM matrix. Secondly, the selection of the optimal fingerprint similarity criterion. The positioning accuracy of the fingerprint matrix with cosine distance is improved by 18.89%, 4.19% and 6.66% respectively compared with the Euclidean distance, chi-square distance and Manhattan distance, and the fingerprint library comparison time is also the shortest, which is shortened by 4.65%, 19.95% and 2.92% respectively. Then, based on the idea of ​​classification, a fingerprint positioning algorithm combining DCNN with five convolutional layers and WKNN matching is proposed to improve the accuracy and robustness of fingerprint positioning.

[0036] The present invention also discloses an electronic device, which includes: a memory for storing a computer program; and a processor for executing the computer program so that the device executes the above method.

[0037] The present invention also provides a computer storage medium, wherein a computer program is stored on the computer-readable storage medium. When the computer program is executed, a device executing the computer program implements the aforementioned method.

[0038] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, and the computer software product can be stored in a storage medium, and the memory can be various types of memory, which can be a random access memory, a read-only memory, a flash memory, etc., such as a read-only memory (English: read-only memory, ROM) / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.

[0039] It should be noted that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely schematic, in which the unit described as a separate component may or may not be physically separated, and the component prompted as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0040] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A fingerprint positioning method based on deep convolutional neural network, characterized in that: The method comprises an offline phase and an online phase; the offline phase comprises the following steps: S1.1, dividing the target positioning area into several sub-areas, and assigning different numbers to each area; S1.

2. Select several reference points in each sub-area, measure and collect the ADCFA matrices of all reference points as fingerprint matrices; construct a data set with the fingerprint matrices, position coordinates and sub-area numbers of all reference points; S1.

3. Build a deep convolutional neural network model and train it using the data set to obtain a trained deep convolutional neural network model; The online phase includes the following steps: S2.1, measure the test points in the target positioning area to obtain a fingerprint matrix; S2.2, input the fingerprint matrix of the test point into the deep convolutional neural network model trained in the offline stage, and output the predicted sub-region number; S2.

3. Calculate the position fingerprints of all reference points in the sub-region corresponding to the sub-region number, calculate the similarity distance between the position fingerprints of all reference points and the position fingerprint of the sub-region number, select several reference points according to the similarity distance, and calculate the specific position coordinates corresponding to the fingerprint of the test point according to the WKNN algorithm to achieve fingerprint positioning.

2. The fingerprint positioning method based on deep convolutional neural network according to claim 1 is characterized in that: The calculation process of the fingerprint matrix includes: S1.2.

1. There is a base station in the target positioning area. The base station obtains the user through channel estimation. The CFR matrix of the signal sent to the base station is recorded as ;user Locate any point in the target area; S1.2.

2. Perform DFT transformation to obtain the angle delay domain channel response matrix, namely the ADCFR matrix, which is recorded as ; S1.2.

3. ADCFR Matrix Calculate the Hadamard inner product with its own conjugate matrix to obtain the angle delay domain channel energy matrix, that is, the ADCFP matrix, recorded as ; S1.2.

4. ADCFP Matrix By taking the square root of all its elements, we can obtain the angle delay domain channel amplitude matrix, namely the ADCFA matrix, which is recorded as .

3. The fingerprint positioning method based on deep convolutional neural network according to claim 2 is characterized in that: The calculation formula include: ， in, for The unitary matrix of is the phase-shift DFT matrix, is the number of antennas of the base station, The rows and columns are , The set of complex matrices of , is the number of subcarriers that send signals to the user; the calculation formulas for the components in the phase-shifted DFT matrix are: ， in, The rows and columns are The square array, For square The element corresponding to the mth row and nth column in .

4. The fingerprint positioning method based on deep convolutional neural network according to claim 2 is characterized in that: The calculation formula include: ， in, is the mean operator, is the number of antennas of the base station, For the ADCFR matrix The conjugate matrix of The rows and columns are The set of real matrices of , for The element corresponding to the mth row and nth column in .

5. According to claim 2, a fingerprint positioning method based on deep convolutional neural network is characterized in that: The calculation formula include: ， in, is the mean operator, is the number of antennas of the base station, For the ADCFP matrix The conjugate matrix of The rows and columns are The set of real matrices of , for The element corresponding to the mth row and nth column in .

6. The fingerprint positioning method based on deep convolutional neural network according to claim 1 is characterized in that: The deep convolutional neural network model includes: The input layer is used to normalize the input data to obtain normalized data. The input data is the fingerprint matrix. Five groups of convolutional layers and pooling layers with the same structure, where: Convolution layer, used to perform convolution operation on the normalized data and output the convolution data; there are 5 convolution layers in total, the first convolution layer has 8 convolution kernels, the second convolution layer has 16 convolution kernels, the third convolution layer has 32 convolution kernels, the fourth convolution layer has 64 convolution kernels, and the fifth convolution layer has 128 convolution kernels; The pooling layer is used to perform maximum pooling on the convolutional data to obtain the pooled data; The fully connected layer is used to perform maximum pooling on the pooled data to obtain the predicted sub-region number.

7. The fingerprint positioning method based on deep convolutional neural network according to claim 1 is characterized in that: The similarity distance is calculated using the cosine distance, and the calculation formula includes: ， in, for arrive The vector of the point, for arrive The vector of the point, for The F-norm of for The F-norm of Represents vectorization; For the base site, , They are test point and reference point respectively.

8. The fingerprint positioning method based on deep convolutional neural network according to claim 1 is characterized in that: The calculation formula of the WKNN algorithm includes: ， in, is the position coordinate obtained using the weighted K nearest neighbor algorithm, is the position coordinate of the kth reference point, is the horizontal and vertical coordinates; is the weight corresponding to the kth reference point and satisfies , is the number of reference points, is the fingerprint distance between the fingerprint matrix at the kth reference point and the fingerprint matrix at the target position to be located, Is a positive number.

9. An electronic device, characterized in that: The device comprises: a memory for storing a computer program; and a processor for executing the computer program so that the device performs a fingerprint positioning method based on a deep convolutional neural network as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that: The computer-readable storage medium stores a computer program. When the computer program is executed, the device executing the computer program implements a fingerprint positioning method based on a deep convolutional neural network as described in any one of claims 1 to 8.

Images