A wireless radio frequency positioning method based on multi-source data and convolutional neural network

Abstract

The application discloses a wireless radio frequency positioning method based on multi-source data and a convolutional neural network, and relates to the field of radio frequency identification positioning. The method comprises the following steps: collecting, by a plurality of fixed antennas, received signal strength indicators and phase angle data of a target label respectively, pre-processing and fusing the collected received signal strength indicators and phase angle data to obtain a two-dimensional tensor; performing preliminary coordinate prediction on the two-dimensional tensor by using a convolutional neural network to obtain preliminary predicted coordinates; the hyperparameters of the convolutional neural network are determined by using an improved multi-role sparrow search algorithm for optimization; the preliminary predicted coordinates are spliced with a plurality of groups of pre-processed received signal strength indicators and phase angle data respectively to obtain a multi-dimensional feature vector, and the multi-dimensional feature vector is input into a multi-layer perceptron for residual correction to obtain predicted three-dimensional coordinates of the target label. The method can improve the positioning accuracy of the target label in an indoor environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radio frequency identification positioning, and particularly relates to a wireless radio frequency positioning method based on multi-source data and a convolutional neural network. BACKGROUND

[0002] With the rapid development of intelligent scenarios, indoor positioning has been widely applied in many fields such as industrial production, warehouse logistics, intelligent buildings, public safety, etc. Due to the factors such as multipath effect, non-line-of-sight (NLOS) in indoor environment, the positioning method based on satellite navigation system is difficult to achieve high-precision positioning in indoor scenarios, and the positioning cost is high.

[0003] Traditional indoor positioning methods include ranging-based positioning technologies such as time of arrival (TOA), time difference of arrival (TDOA), angle of arrival (AOA), which usually rely on special hardware devices, and the system deployment is complex, and the anti-interference ability is limited in complex environment. In recent years, the development of deep learning technology provides a new solution for indoor positioning, and some research has tried to apply convolutional neural network to coordinate estimation. However, the precision of the existing convolutional neural network for indoor target coordinate estimation is still low. SUMMARY

[0004] Therefore, it is necessary to provide a wireless radio frequency positioning method based on multi-source data and a convolutional neural network to solve the above technical problems. The method improves the positioning precision of the target tag in the indoor environment.

[0005] The present application adopts the following technical solutions:

[0006] The present application provides a wireless radio frequency positioning method based on multi-source data and a convolutional neural network, comprising:

[0007] The received signal strength indicators and phase angle data of the target tag are collected by multiple fixed antennas respectively, and the collected received signal strength indicators and phase angle data are preprocessed and fused to obtain a two-dimensional tensor; the multiple fixed antennas are arranged at different positions respectively;

[0008] The two-dimensional tensor is preliminarily predicted by the convolutional neural network to obtain a preliminary prediction coordinate;

[0009] The preliminary prediction coordinate is spliced with the multiple groups of preprocessed received signal strength indicators and phase angle data respectively to obtain a multi-dimensional feature vector, and the multi-dimensional feature vector is input into a multi-layer perceptron for residual correction to obtain a predicted three-dimensional coordinate of the target tag.

[0010] Preferably, during the construction of the convolutional neural network, an improved multi-role sparrow search algorithm is used to determine the hyperparameters, specifically including:

[0011] Initialize the parameters of the improved multi-role sparrow search algorithm; the parameters include the maximum number of iterations, population size, hyperparameter search range, and role ratio;

[0012] The fitness value is calculated using the average Euclidean error between the predicted coordinates of the convolutional neural network and the true coordinates of the sample two-dimensional tensor data as the fitness function.

[0013] The positions of different characters are updated and the proportions of different characters are adjusted based on fitness values;

[0014] Based on the number of iterations, the update magnitude of different roles is adjusted through an adaptive step size strategy;

[0015] After updating the positions of different characters, adjusting the proportion of different characters, and adjusting the update range of different characters, recalculate the fitness value according to the fitness function, and record the optimal character and the error of the optimal character after updating the fitness value.

[0016] Determine whether the maximum number of iterations has been reached, and determine the hyperparameters in the position of the optimal role when the maximum number of iterations has been reached as the optimized hyperparameters; the hyperparameters include the learning rate, the number of first convolutional kernels, the number of second convolutional kernels, and the number of neurons in the first fully connected layer.

[0017] Preferably, the multiple roles include producers, followers, scavengers, explorers, and jumpers; producers are used to perform local searches, followers are used to move closer to the optimal solution, scavengers are used to perform perturbation searches based on the optimal solution, explorers are used to perform global walks, and jumpers are used to mutate the dimensions in the position; the producer's position update formula is:

[0018] ;

[0019] in, The reduction factor is... Indicates the first i The producer in the first t The first generation j The values ​​of each hyperparameter, step For adaptive step size, Indicates the first i The producer in the first t+ The first generation j The values ​​of each hyperparameter;

[0020] The formula for updating the position of followers is:

[0021] ;

[0022] where, denotes the position vector of the t th follower in the i th generation, denotes the position of the best follower in the current population, is the weight coefficient, denotes the position vector of the t+ th follower in the i th generation;

[0023] The position update formula of the scavenger is:

[0024] ;

[0025] where, is the normal perturbation variable, denotes the value of the t th hyperparameter of the current global optimal solution in the j th generation, is the perturbation factor, is the value of the i th hyperparameter of the t+ th scavenger in the j th generation;

[0026] The position update formula of the explorer is:

[0027] ;

[0028] where, denotes the position vector of the t+ th explorer in the i th generation, is the standard normal distribution, denotes the position vector of the t th explorer in the i th generation, step is the adaptive step size;

[0029] The position update formula of the jumper is:

[0030] ;

[0031] where, denotes the position vector of the i th jumper in the d th dimension, denotes the random sampling in the d th dimension of the search space.

[0032] Preferably, the multilayer perceptron includes three independent error regression networks; the three independent error regression networks are a first error regression network, a second error regression network, and a third error regression network; each error regression network includes two cascaded hidden layers; the multidimensional feature vector is input into the multilayer perceptron for residual correction to obtain the predicted three-dimensional coordinates of the target label, specifically including:

[0033] In the multidimensional feature vector X The coordinates are input into the first error regression network, which then inputs the multidimensional feature vectors into the network. Y The coordinates are input into the second error regression network, which then inputs the multidimensional feature vectors into the network. Z The coordinates are input into the third error regression network;

[0034] The first error regression network uses two cascaded hidden layers to analyze the multidimensional feature vector. X The coordinates are subjected to error regression to obtain... X Residual values ​​in direction;

[0035] The second error regression network uses two cascaded hidden layers to analyze the multidimensional feature vectors. Y The coordinates are subjected to error regression to obtain... Y Residual values ​​in direction;

[0036] The multidimensional feature vector is processed through two cascaded hidden layers in the third error regression network. Z The coordinates are subjected to error regression to obtain... Z Residual values ​​in direction;

[0037] Will X Predicted values ​​of coordinates and X The residual values ​​in the directions are added together to obtain X Coordinates, Y Predicted values ​​of coordinates and Y The residual values ​​in the directions are added together to obtain Y Coordinates, Z Predicted values ​​of coordinates and Z The residual values ​​in the directions are added together to obtain Z coordinate;

[0038] Will X coordinate, Y coordinates and Z The coordinates are determined as the predicted three-dimensional coordinates of the target label.

[0039] Preferably, the acquired received signal strength index and phase angle data are preprocessed and fused to obtain a two-dimensional tensor, specifically including:

[0040] The received signal strength and phase angle data of the target tag collected by each fixed antenna are filtered.

[0041] The missing value interpolation completion is performed on the received signal strength indicator and phase angle data after the filtering processing, and the received signal strength indicator and phase angle data after the missing value interpolation completion are fused according to a fusion weight to obtain a fusion feature vector;

[0042] The plurality of fusion feature vectors of the target label are constructed into a two-dimensional tensor; the two-dimensional tensor is C × L ; L N is the number of antennas, C is the channel number of the received signal strength indicator and phase angle data after the missing value interpolation completion.

[0043] Preferably, the missing value interpolation completion is performed on the received signal strength indicator and phase angle data after the filtering processing, and specifically includes:

[0044] The received signal strength indicator and phase angle data after the filtering processing are completed through linear interpolation completion;

[0045] When the calculation result corresponding to the linear interpolation completion is empty, the received signal strength indicator and phase angle data after the preprocessing are completed through the nearest neighbor method.

[0046] Preferably, the filtering processing includes one-dimensional Gaussian filter smoothing processing and one-dimensional Kalman filter smoothing processing.

[0047] The application provides a wireless radio frequency positioning device based on multi-source data and a convolutional neural network, comprising:

[0048] A preprocessing module is configured to collect received signal strength indicators and phase angle data of a target label through a plurality of fixed antennas respectively, and to preprocess and fuse the collected received signal strength indicators and phase angle data to obtain a two-dimensional tensor; the plurality of fixed antennas are arranged at different positions respectively.

[0049] A prediction module is configured to perform preliminary coordinate prediction on the two-dimensional tensor through a convolutional neural network to obtain preliminary predicted coordinates.

[0050] A positioning module is configured to splice the preliminary predicted coordinates with a plurality of groups of received signal strength indicators and phase angle data after preprocessing respectively to obtain a multi-dimensional feature vector, and to input the multi-dimensional feature vector into a multi-layer perception machine for residual correction to obtain a predicted three-dimensional coordinate of the target label.

[0051] The application provides a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is executed by a processor to implement the wireless radio frequency positioning method based on multi-source data and a convolutional neural network.

[0052] The application provides a computer device, comprising a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the processor implements the above-mentioned wireless radio frequency positioning method based on multi-source data and a convolutional neural network when executing the program.

[0053] The application adopts the above-mentioned at least one technical scheme to achieve the following beneficial effects:

[0054] The received signal strength index and phase angle data collected by the multiple fixed antennas are preprocessed and fused to obtain a two-dimensional tensor, multi-dimensional observation data of the received signal strength index signal strength and phase angle are obtained through the distributed antenna array, the defects that single-point measurement is susceptible to multipath effect are effectively overcome, preprocessing operation is performed on the collected signals to eliminate environmental interference and improve the data signal-to-noise ratio, the two-dimensional tensor is preliminarily coordinate-predicted through the convolutional neural network to obtain a preliminary prediction coordinate, the preliminary prediction coordinate is spliced with multiple groups of the received signal strength index and phase angle data after preprocessing to obtain a multi-dimensional feature vector, and the multi-dimensional feature vector is input into a multi-layer perceptron for residual correction to obtain a predicted three-dimensional coordinate of the target label, and the linear deviation of the convolutional neural network is compensated through the multi-layer perceptron. The method improves the positioning accuracy of the target label in the indoor environment.

[0055] In addition, the hyperparameters of the convolutional neural network are determined by optimization through the improved multi-role sparrow search algorithm, the search efficiency of the hyperparameters is improved, and the training of the convolutional neural network is effectively prevented from falling into local optimization, and thus the prediction accuracy of the convolutional neural network is improved. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings, which are included to provide a further understanding of the application, constitute a part of this application, and the illustrative embodiments of the application and their description serve to explain the application, and do not limit the application in any way. In the drawings:

[0057] Figure 1 A wireless radio frequency positioning method based on multi-source data and a convolutional neural network is provided.

[0058] Figure 2 A multi-role sparrow search algorithm for optimizing hyperparameters is provided.

[0059] Figure 3 A convolutional neural network is provided.

[0060] Figure 4 A multi-layer perceptron working process is provided.

[0061] Figure 5A wireless radio frequency positioning block diagram based on multi-source data and a convolutional neural network is provided for the present application.

[0062] Figure 6 A wireless radio frequency positioning device schematic diagram based on multi-source data and a convolutional neural network is provided for the present application.

[0063] Figure 7 A computer device schematic diagram for implementing a wireless radio frequency positioning method based on multi-source data and a convolutional neural network is provided for the present application. DETAILED DESCRIPTION

[0064] To make the objectives, technical solutions, and advantages of the present application clearer, the technical solutions of the present application will be described below in conjunction with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0065] Devices such as desktop computers, servers, notebook computers, etc. that can execute the solutions of the present application. For the convenience of explanation, only servers will be described below as the execution subject.

[0066] Due to factors such as multipath effect, non-line-of-sight (NLOS) propagation, etc. in indoor environments, traditional satellite navigation system-based positioning methods, such as satellite navigation system-based positioning methods, usually rely on special hardware devices, system deployment is complex, and anti-interference capability is limited in complex environments; for example, non-rangefinding-based positioning methods construct a fingerprint database by collecting wireless signal information such as received signal strength indicator (RSSI), etc. and perform positioning through online matching. However, such methods are sensitive to environmental changes, have poor signal stability, and the fingerprint data itself is easily disturbed by noise, affecting positioning accuracy.

[0067] In recent years, the development of deep learning technology has provided a new solution for indoor positioning. Convolutional neural networks have strong non-linear modeling and feature extraction capabilities, and some research has attempted to apply convolutional neural networks to the processing and analysis of RSSI, phase angle, and other signal data to improve the robustness and accuracy of positioning systems. However, existing solutions mostly apply convolutional neural networks to position classification tasks or preliminary coordinate estimation, and the selection of parameters relies on human experience, making it difficult to achieve fast and effective optimal parameter determination.

[0068] Existing methods mostly rely on a single type of feature, lacking in-depth utilization of multi-source feature fusion strategies. In actual deployment, due to environmental interference, uneven antenna layout, or equipment errors, even if the deep learning model has certain prediction ability, the output coordinates may still have systematic deviations, making it difficult to meet the high-precision positioning requirements in complex scenarios.

[0069] The prior art needs an indoor electronic tag positioning method that can fuse multi-source positioning features, adaptively adjust model parameters combined with intelligent optimization algorithms, and has an error correction mechanism to improve the overall positioning accuracy and environmental adaptability of the system.

[0070] The technical solutions provided by the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0071] Figure 1 The flowchart of the wireless radio frequency positioning method based on multi-source data and a convolutional neural network in the present application specifically includes the following steps:

[0072] S101: Collect the received signal strength indicators and phase angle data of the target tag through multiple fixed antennas, preprocess and fuse the collected received signal strength indicators and phase angle data, and obtain a two-dimensional tensor; the multiple fixed antennas are arranged at different positions.

[0073] Multiple fixed antennas are arranged in a certain experimental space, such as at coordinate points (X1, Y1, Z1), (X2, Y2, Z2), and (X3, Y3, Z3), etc. RSSI and phase angle data of the tag at different positions are collected using multiple antennas.

[0074] In an exemplary embodiment, the filtering process includes one-dimensional Gaussian filter smoothing and one-dimensional Kalman filter smoothing.

[0075] Specifically, to reduce the influence of environmental noise on the signal, first apply Gaussian filtering to each set of RSSI and phase angle data for preliminary smoothing, and then apply a Kalman filter to further remove residual noise, improving data stability and continuity.

[0076] In an exemplary embodiment, the collected received signal strength indicators and phase angle data are preprocessed and fused to obtain a two-dimensional tensor, specifically including: filtering the received signal strength indicators and phase angle data of the target tag collected by each fixed antenna; performing interpolation and completion of missing values on the received signal strength indicators and phase angle data after filtering, and fusing the received signal strength indicators and phase angle data after interpolation and completion of missing values according to the fusion weight to obtain a fusion feature vector; constructing multiple fusion feature vectors of the target tag into a two-dimensional tensor; the two-dimensional tensor is C xL ; L is the number of antennas, C is the channel number of the received signal strength indicator and phase angle data after interpolation completion of missing values.

[0077] In an exemplary embodiment, the received signal strength indicator and phase angle data after filtering are interpolated and completed, specifically including: completing the received signal strength indicator and phase angle data after filtering by linear interpolation completion; when the calculation result corresponding to the linear interpolation completion is empty, completing the received signal strength indicator and phase angle data after preprocessing by the nearest neighbor method.

[0078] Specifically, due to the missing of part of the data disturbed by environmental factors in the measurement process, linear interpolation completion is needed for each antenna data; if the interpolation fails, the nearest neighbor method is used to complete the missing data. The completed RSSI and phase angle are standardized and weighted average fused according to a certain fusion weight to form a fusion feature vector. Linear interpolation and nearest neighbor completion strategy are used for missing data to ensure data integrity.

[0079] For each coordinate point, the fusion feature data of the antennas are arranged in order to construct a two-dimensional tensor like C × L (wherein, C is the number of phase angle channels, L is the number of antennas).

[0080] Specifically, the preprocessed RSSI and phase angle are linearly fused according to a certain fusion weight to obtain a fusion feature vector, and are combined into a two-dimensional tensor in order of multiple antennas; each sample and its corresponding true three-dimensional coordinates form a sample pair, which is divided into a training set and a validation set according to a certain proportion. The fusion feature vector and its corresponding true coordinates are bound as a sample pair, and all samples are divided into a training set and a validation set according to a certain proportion. This data will be used as input for subsequent deep model training and evaluation.

[0081] S102: performing preliminary coordinate prediction on the two-dimensional tensor by a convolutional neural network to obtain preliminary predicted coordinates; the hyperparameters of the convolutional neural network are determined by an improved multi-role sparrow search algorithm; the multi-roles include producers, followers, scavengers, explorers and jumpers; the producers are used for performing local search, the followers are used for approaching the optimal solution, the scavengers are used for disturbance search based on the optimal solution, the explorers are used for global wandering, and the jumpers are used for implementing mutation on the dimensions in the position.

[0082] In an exemplary embodiment, the improved multi-role sparrow search algorithm is used to determine the hyperparameters of the convolutional neural network, specifically including: initializing the parameters of the improved multi-role sparrow search algorithm; the parameters include the maximum number of iterations, the population size, the hyperparameter search range and the role ratio; taking the average Euclidean error of the predicted coordinates of the convolutional neural network and the true coordinates of the sample two-dimensional tensor data as the fitness function, and calculating the fitness value; updating the positions of different roles based on the fitness value and adjusting the number ratio of different roles; based on the number of iterations, adjusting the update amplitude of different roles through a step adaptive strategy; after updating the positions of different roles, adjusting the number ratio of different roles and adjusting the update amplitude of different roles, the fitness value is recalculated according to the fitness function, and the optimal role after updating the fitness value and the error of the optimal role are recorded; judge whether the maximum number of iterations is reached, and the hyperparameters in the position of the optimal role when the maximum number of iterations is reached are determined as the optimized hyperparameters; the hyperparameters include the learning rate, the first convolution kernel number, the second convolution kernel number and the first fully connected layer neuron number.

[0083] Figure 2 The multi-role sparrow search algorithm provided by the application is shown in the schematic diagram of optimizing hyperparameters, as shown in Figure 2 The multi-role sparrow search algorithm includes the following sub-processes:

[0084] Step one: initialize the search range of the convolutional neural network (CNN) hyperparameters, set the algorithm parameters and the multi-role ratio. Set the population size, the maximum number of iterations, initialize the population individuals, and divide the individuals into multiple roles such as producers, followers, scavengers, explorers and jumpers. The producer P is used for local search, the follower F is used for approaching the optimal solution, the scavenger S is used for disturbance search based on the optimal solution, the explorer E is used for global wandering, and the jumper J is used for implementing mutation on the dimensions in the position.

[0085] Step two, calculate the fitness of each individual and record the current optimal individual.

[0086] Step three: update the position according to the role, and adjust the update amplitude of different roles using a step adaptive strategy.

[0087] Step adaptive strategy: the step of each iteration is updated with the iteration increasing, corresponding to formula (1):

[0088] (1)

[0089] Wherein, step is the adaptive step, which increases linearly with the iteration round, t is the current generation, T is the maximum number of iterations.

[0090] The fitness of individuals in the initial population is evaluated, and they are sorted according to their fitness values. The current global best individual and its error value are recorded. The fitness function (validation set average Euclidean error) is calculated as shown in formula (2):

[0091] (2);

[0092] in, N This represents the total number of sample points in the validation set. x i , y i , z i Indicates the first i Predicted coordinates of each sample point , , Indicates the first i The true coordinates of each sample point.

[0093] The update behavior for each character is implemented using the following formula.

[0094] The producer's position update formula is shown in formula (3):

[0095] (3);

[0096] in, =0.1 is the reduction factor. Indicates the first i The producer in the first t The first generation j The values ​​of each hyperparameter, Indicates the first i The producer in the first t+ The first generation j The values ​​of each hyperparameter.

[0097] The position update formula for followers is shown in formula (4):

[0098] (4);

[0099] Where, in the formula Indicates the first t The middle generation i The position vectors of the followers This indicates the position of the best follower in the current population. These are the weighting coefficients. , Indicates the first t+ 1st generation i The position vectors of the followers.

[0100] The position update formula of the scavenger is shown as formula (5):

[0101] (5);

[0102] wherein, is a normal disturbance variable, , represents the value in the first t dimension in the first j generation, is a disturbance factor, and the default is step , represents the value of the first i scavenger in the first t+ generation in the first j hyperparameter.

[0103] The position update formula of the explorer is shown as formula (6):

[0104] (6);

[0105] wherein, represents the position vector of the first t+ explorer in the first i generation, is a standard normal distribution.

[0106] The position update formula of the jumper is shown as formula (7):

[0107] (7);

[0108] wherein, represents the position vector of the first i jumper in the first d dimension, represents a random sampling in the first d dimension of the search space.

[0109] Step four: the fitness of the current population individuals is evaluated, and it is judged whether the maximum iteration number is reached or the early termination condition is met.

[0110] Step five: the current optimal hyperparameter combination and the error value of the corresponding convolutional neural network on the verification set are output.

[0111] Figure 3 The model structure diagram of the convolutional neural network model provided in the application is shown as shown in Figure 3 The improved multi-role sparrow search algorithm is used for hyperparameter optimization of the convolutional neural network, and the optimal hyperparameter combination obtained includes a learning rate l r , a first convolution kernel numberc 1, the number of second convolution kernels c 2, and the number of first fully connected layer neurons f 1. The convolutional neural network comprises a convolutional layer 1, a convolutional layer 2, a flattening layer, a first fully connected layer, and a second fully connected layer. The input channel of the convolutional layer 1 is 2, the output channel is c 1, the convolution kernel size is 2, the step is 1, and the activation function is ReLU. The input channel of the convolutional layer 2 is c 1, the output channel is c 2, the convolution kernel size is 2, the step is 1, and the activation function is ReLU. The output dimension of the flattening layer is c 2, the input dimension of the first fully connected layer is c 2, the output dimension is f 1, the activation function is ReLU, and the output of the second fully connected layer is the preliminary predicted coordinates.

[0112] The convolutional neural network model is fully trained on the training set, and the loss function in the training process adopts the Euclidean distance between the predicted coordinates and the real labels. After the training is completed, the model is used to predict the validation set data, and the preliminary predicted coordinates of the indoor target are output.

[0113] S103: The preliminary predicted coordinates are respectively spliced with a plurality of groups of preprocessed received signal strength indicators and phase angle data to obtain a multi-dimensional feature vector, and the multi-dimensional feature vector is input into a multi-layer perceptron for residual correction to obtain a predicted three-dimensional coordinate of the target label.

[0114] In an exemplary embodiment, the multi-layer perceptron comprises three independent error regression networks; the three independent error regression networks are a first error regression network, a second error regression network, and a third error regression network; each error regression network comprises two serially connected hidden layers; the multi-dimensional feature vector is input into the multi-layer perceptron for residual correction to obtain a predicted three-dimensional coordinate of the target label, specifically comprising: inputting the X coordinate in the multi-dimensional feature vector into the first error regression network, inputting the Y coordinate in the multi-dimensional feature vector into the second error regression network, and inputting the Z coordinate in the multi-dimensional feature vector into the third error regression network; performing error regression operation on the X coordinate in the multi-dimensional feature vector through the two serially connected hidden layers in the first error regression network to obtain a residual value of the X direction; performing error regression operation on the Y coordinate in the multi-dimensional feature vector through the two serially connected hidden layers in the second error regression network to obtain a residual value of the Y direction; and performing error regression operation on the Z coordinate in the multi-dimensional feature vector through the two serially connected hidden layers in the third error regression network to obtain a residual value of theZ the residual value of the direction; adding X the predicted value of the coordinate and X the residual value of the direction to obtain X the coordinate, adding Y the predicted value of the coordinate and Y the residual value of the direction to obtain Y the coordinate, adding Z the predicted value of the coordinate and Z the residual value of the direction to obtain Z the coordinate; adding X the coordinate, Y the coordinate and Z the coordinate is determined as the predicted three-dimensional coordinate of the target tag.

[0115] Specifically, the predicted three-dimensional coordinate predicted by the convolutional neural network is expanded and spliced with the original input fusion feature to form a multi-dimensional feature vector, which is respectively input to three independent multi-layer perceptrons (MLP), which are respectively used for regression of the residual values of the X, Y and Z directions. Each MLP structure includes an input layer, a hidden layer 1, a hidden layer 2 and an output layer. The residual vector output by the output layer is added to the predicted coordinates of the convolutional neural network item by item to obtain the final corrected three-dimensional positioning result. The input layer is a multi-dimensional feature vector, the input is the RSSI and phase angle data of the CNN predicted coordinates and multiple antennas, the hidden layer 1 is a fully connected layer, the input dimension is H1, the output dimension is H1, and the activation function is ReLU. The hidden layer 2 is a fully connected layer, the input dimension is H1, the output dimension is H2, and the activation function is ReLU. The output dimension of the output layer is 1. n n

[0116] Figure 4 The working process diagram of the multi-layer perceptron provided by the application is as shown in Figure 4 , the CNN predicted coordinates in Figure 4 , the predicted three-dimensional coordinates, Figure 4 , the preprocessed multiple antenna RSSI and phase angle values, i.e., the multiple preprocessed received signal strength indicators and phase angle data, are spliced by a feature splicing module to obtain a multi-dimensional feature vector. The multi-dimensional feature vector is respectively input to MLP-ΔX, MLP-ΔY and MLP-ΔZ, wherein MLP-ΔX is a first error regression network, MLP-ΔY is a second error regression network, and MLP-ΔZ is a third error regression network. In MLP-ΔX, the residual value of the X direction is predicted through the hidden layer 1 and the hidden layer 2, in MLP-ΔY, the residual value of the Y direction is predicted through the hidden layer 1 and the hidden layer 2, and in MLP-ΔZ, the residual value of the Z direction is predicted through the hidden layer 1 and the hidden layer 2. X Y ​​​The residual values ​​in the Z direction are predicted in MLP-ΔZ using hidden layers 1 and 2. X Predicted values ​​of coordinates and X The residual values ​​in the directions are added together to obtain X Coordinates, Y Predicted values ​​of coordinates and Y The residual values ​​in the directions are added together to obtain Y Coordinates, Z Predicted values ​​of coordinates and Z The residual values ​​in the directions are added together to obtain Z Coordinates; X coordinate, Y coordinates and Z The coordinates are determined as the predicted three-dimensional coordinates of the target label.

[0117] In one exemplary implementation, the present invention provides as follows: Figure 5 The block diagram shown is for a radio frequency localization method based on multi-source data and convolutional neural networks, as follows: Figure 5 As shown, the data acquisition module is used to acquire the RSSI and phase angle of the target tag through multiple fixed antennas. The data preprocessing module performs Gaussian Kalman filtering, linear interpolation, and data augmentation on the data acquired by the data acquisition module, and then inputs it into the multi-source feature fusion module to generate multi-dimensional feature vectors. The multi-sparrow search algorithm optimization module optimizes the hyperparameters of the CNN to obtain a trained CNN. The trained CNN is used to predict the preliminary predicted coordinates of the target tag. The MLP error correction module corrects the preliminary predicted coordinates to obtain the predicted three-dimensional coordinates of the target tag.

[0118] Experimental verification and performance analysis. Using the same validation set data, the positioning accuracy of the unoptimized convolutional neural network, the multi-role sparrow search algorithm-optimized CNN, and both after introducing MLP residual correction were compared. The results show that the average positioning error of the proposed method on the validation set decreased from the initial 0.2607 m to 0.0566 m, a reduction of more than 78%, verifying the accuracy advantage of the proposed method in complex indoor environments.

[0119] When applying the radio frequency positioning method based on multi-source data and convolutional neural networks provided by this invention, it is not necessary to... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this invention does not impose any restrictions on it.

[0120] The above describes a wireless radio frequency positioning method based on multi-source data and convolutional neural networks, provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding wireless radio frequency positioning device based on multi-source data and convolutional neural networks, such as... Figure 6 As shown.

[0121] Figure 6 A wireless radio frequency positioning device based on multi-source data and a convolutional neural network provided by the application is shown in the schematic diagram, comprising:

[0122] The preprocessing module 601 is used for collecting the received signal strength indicators and phase angle data of the target label through multiple fixed antennas respectively, and pre-processing and fusing the collected received signal strength indicators and phase angle data to obtain a two-dimensional tensor; the multiple fixed antennas are arranged at different positions respectively.

[0123] The prediction module 602 is used for performing preliminary coordinate prediction on the two-dimensional tensor through a convolutional neural network to obtain preliminary predicted coordinates.

[0124] The positioning module 603 is used for splicing the preliminary predicted coordinates with multiple groups of received signal strength indicators and phase angle data after preprocessing respectively to obtain a multi-dimensional feature vector, and inputting the multi-dimensional feature vector into a multi-layer perception machine for residual correction to obtain the predicted three-dimensional coordinates of the target label.

[0125] The specific limitations of the wireless radio frequency positioning device based on multi-source data and a convolutional neural network can be referred to the limitations of the wireless radio frequency positioning method based on multi-source data and a convolutional neural network in the foregoing, which will not be repeated here. Each module in the wireless radio frequency positioning device based on multi-source data and a convolutional neural network described above can be realized by software, hardware and combinations thereof in whole or in part. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to call and execute the operations corresponding to each module by the processor.

[0126] The application further provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the wireless radio frequency positioning method based on multi-source data and a convolutional neural network. Figure 1 The application provides a wireless radio frequency positioning method based on multi-source data and a convolutional neural network.

[0127] The application further provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the wireless radio frequency positioning method based on multi-source data and a convolutional neural network. Figure 7 The structure schematic diagram of the computer device is shown in the figure, and the computer device includes a processor, an internal bus, a network interface, a memory and a non-volatile memory at the hardware level, and of course can also include other hardware required by the business. Figure 7 The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs to realize the wireless radio frequency positioning method based on multi-source data and a convolutional neural network provided by the application. Figure 1 The application provides a wireless radio frequency positioning method based on multi-source data and a convolutional neural network.

[0128] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments of the methods. In the embodiments of the present application, any reference to memory, storage, database or other medium can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0129] The technical features of the above embodiments can be combined in any way. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

Claims

1. A wireless radio frequency positioning method based on multi-source data and convolutional neural networks, characterized in that, include: The received signal strength index and phase angle data of the target tag are collected by multiple fixed antennas. The collected received signal strength index and phase angle data are preprocessed and fused to obtain a two-dimensional tensor. The multiple fixed antennas are arranged at different positions. Preliminary coordinate predictions are obtained by performing initial coordinate predictions on the two-dimensional tensor using a convolutional neural network (CNN). During the construction of the CNN, an improved multi-role sparrow search algorithm is used to determine hyperparameters. Specifically, this includes: initializing the parameters of the improved multi-role sparrow search algorithm; these parameters include the maximum number of iterations, population size, hyperparameter search range, and role ratio; calculating the fitness value using the average Euclidean error between the predicted coordinates of the CNN and the true coordinates of the sample two-dimensional tensor data as the fitness function; updating the positions of different roles and adjusting the ratio of different roles based on the fitness value; adjusting the update magnitude of different roles using a step-size adaptive strategy based on the number of iterations; and updating the positions of different roles and adjusting the ratio of different roles. After adjusting the update magnitude of different roles, the fitness value is recalculated according to the fitness function, and the optimal role and its error are recorded after updating the fitness value. It is determined whether the maximum number of iterations has been reached, and the hyperparameters in the position of the optimal role corresponding to the maximum number of iterations are determined as the optimized hyperparameters. The hyperparameters include the learning rate, the number of first convolutional kernels, the number of second convolutional kernels, and the number of neurons in the first fully connected layer. The multiple roles include producers, followers, scavengers, explorers, and jumpers. The producers are used to perform local search, the followers are used to move closer to the optimal solution, the scavengers are used to perform perturbation search based on the optimal solution, the explorers are used to perform global walks, and the jumpers are used to mutate the dimensions in the position. The preliminary predicted coordinates are concatenated with multiple sets of preprocessed received signal strength indicators and phase angle data to obtain a multidimensional feature vector. The multidimensional feature vector is then input into a multilayer perceptron for residual correction to obtain the predicted three-dimensional coordinates of the target label.

2. The method as described in claim 1, characterized in that, The producer's position update formula is: ； in, The reduction factor is... Indicates the first i The producer in the first t The first generation j The values ​​of each hyperparameter, step For adaptive step size, Indicates the first i The producer in the first t+ The first generation j The values ​​of each hyperparameter; The formula for updating the position of followers is: ； in, Indicates the first t The middle generation i The position vectors of the followers This indicates the position of the best follower in the current population. These are the weighting coefficients. Indicates the first t+ 1st generation i The position vectors of the followers; The formula for updating the position of scavengers is: ； in, For normally distributed perturbation variables, Indicates the first t The current global optimal solution in the generation is at the th j The values ​​of each hyperparameter, As a disturbance factor, For the first i The scavenger in the first t+ The first generation j The values ​​of each hyperparameter; The explorer's position update formula is: ； in, Indicates the first t+ 1st generation i The position vectors of the explorers It follows a standard normal distribution. Indicates the first t The middle generation i The position vectors of the explorers step For adaptive step size; The jumper's position update formula is: ； in, Indicates the first i The jumper in the first d A position vector with 3 dimensions. Indicates the number of search spaces d Random sampling of dimensions.

3. The method as described in claim 1, characterized in that, The multilayer perceptron includes three independent error regression networks; these three independent error regression networks are a first error regression network, a second error regression network, and a third error regression network; each error regression network includes two cascaded hidden layers; the step of inputting the multidimensional feature vector into the multilayer perceptron for residual correction to obtain the predicted three-dimensional coordinates of the target label specifically includes: In the multidimensional feature vector X The coordinates are input into the first error regression network, and the multidimensional feature vectors are... Y The coordinates are input into the second error regression network, and the multidimensional feature vectors are... Z The coordinates are input into the third error regression network; The first error regression network uses two cascaded hidden layers to analyze the multidimensional feature vector. X The coordinates are subjected to error regression to obtain... X Residual values ​​in direction; The second error regression network uses two cascaded hidden layers to analyze the multidimensional feature vector. Y The coordinates are subjected to error regression to obtain... Y Residual values ​​in direction; The third error regression network uses two cascaded hidden layers to process the multidimensional feature vector. Z The coordinates are subjected to error regression to obtain... Z Residual values ​​in direction; Will X Predicted values ​​of coordinates and X The residual values ​​in the directions are added together to obtain X Coordinates, Y Predicted values ​​of coordinates and Y The residual values ​​in the directions are added together to obtain Y Coordinates, Z Predicted values ​​of coordinates and Z The residual values ​​in the directions are added together to obtain Z coordinate; The X Coordinates, as stated Y Coordinates and the stated Z The coordinates are determined as the predicted three-dimensional coordinates of the target label.

4. The method as described in claim 1, characterized in that, The acquired received signal strength and phase angle data are preprocessed and fused to obtain a two-dimensional tensor, specifically including: The received signal strength and phase angle data of the target tag collected by each fixed antenna are filtered. The received signal strength index and phase angle data after filtering are filled with missing values ​​through interpolation, and the received signal strength index and phase angle data after missing value interpolation are fused with the fusion weight to obtain the fused feature vector. The multiple fused feature vectors of the target label are constructed into the two-dimensional tensor; the two-dimensional tensor is... C × L The L The number of antennas, the C The number of channels for the received signal strength index and phase angle data after interpolation to complete the missing values.

5. The method as described in claim 4, characterized in that, The process of interpolating and filling in missing values ​​in the received signal strength index and phase angle data after filtering specifically includes: Linear interpolation is used to complete the received signal strength and phase angle data after filtering. When the calculation result corresponding to the linear interpolation completion is empty, the preprocessed received signal strength index and phase angle data are completed using the nearest neighbor method.

6. The method as described in claim 5, characterized in that, The filtering process includes one-dimensional Gaussian filter smoothing and one-dimensional Kalman filter smoothing.

Images