Aerial radio map positioning method and system based on deep learning

Abstract

The present application discloses an aerial radio map positioning method based on deep learning and its system. The aerial radio map positioning method based on deep learning specifically includes the following steps: collecting radio spectrum data; processing the collected radio spectrum data; building a neural network model based on the processed radio spectrum data; in response to the completion of the neural network model, inputting the processed radio spectrum data into the built neural network model to train the neural network model; in response to completing a round of neural network model training, performing neural network model performance evaluation; judging whether the accuracy requirements of the radio positioning system are met; if the accuracy requirements of the radio positioning system are met, saving the neural network model and applying it. The radio positioning system based on data training in the present application no longer needs to understand the information of the radio propagation environment, the transmitter and receiver of the radio signal, and is more universal.

Description

Technical Field

[0001] The present application relates to the field of communications, and in particular, to an aerial radio map positioning method and system based on deep learning. Background Art

[0002] In recent years, with the rapid development of global science and technology, the continuous popularization of mobile devices and the large-scale deployment of wireless networks, wireless technology has had varying degrees of impact on all aspects of human beings. Due to the practical value of wireless technology, radio map-based positioning services have become the focus of research in the industry and community, that is, accurately positioning one's own position based on the radio spectrum data obtained by the user. At present, the existing geometric positioning methods mainly use information such as the angle between the signal transmitter and the signal receiver, the time between the signal transmitted by the signal transmitter and the receiver to calculate the distance and angle between the two, so as to achieve the positioning needs. However, since the real environment is often blocked by obstacles (such as high-rise buildings, trees, etc.), the signal received by the signal receiver will be affected by the propagation environment, and the transmission speed of the radio signal is close to the speed of light, so the positioning using the receiving time and angle is inaccurate and may have large errors. In addition, the existing geometric method can also be positioned according to the strength of the received signal. Since the signal has the characteristic of attenuation during the propagation of space, the distance and relative position between the transmitting signal source and the receiving signal source are estimated according to the degree of attenuation of the wireless signal during the propagation process and the attenuation formula. However, the attenuation characteristics of wireless signals in real environments are difficult to estimate. Even if the attenuation characteristics are estimated, the error is very large, resulting in inaccurate positioning. Although the deep learning-based aerial radio map positioning method is no longer affected by traditional positioning technology and provides a better method for users' precise positioning, it still has many difficult problems to solve. These mainly include:

[0003] 1) A large amount of accurate radio spectrum data is needed to support the training of neural networks, because radio positioning based on deep learning is data-based. When the selected algorithm is suitable for the neural network, the more data collected, the stronger the generalization ability of the network model of the final neural network, and the more accurate the prediction. However, the collection of a large amount of data will increase the cost of the positioning system. How to use limited radio spectrum data to design an accurate radio positioning system is an urgent problem that needs to be solved. 2) Deep learning neural networks have a great impact on training accuracy. For the same data, different neural networks (number of network layers, hyperparameters, etc.) are different, and the final neural network generalization model will also be different. Therefore, how to reasonably design a neural network to improve the accuracy of the final radio positioning is also a problem that needs to be solved.

[0004] Therefore, how to use limited radio data through deep learning to achieve high-precision aerial positioning is a problem that technicians in this field urgently need to solve. Summary of the invention

[0005] The present application provides an aerial radio map positioning method based on deep learning, which specifically includes the following steps: S1, collecting radio spectrum data; S2, processing the collected radio spectrum data; S3, building a neural network model based on the processed radio spectrum data; S4, in response to the completion of the construction of the neural network model, inputting the processed radio spectrum data into the built neural network model to train the neural network model; S5, in response to completing a round of neural network model training, performing neural network model performance evaluation; S6, if the result of the neural network model performance evaluation is passed, judging whether the radio positioning system accuracy requirement is met; S7, if the radio positioning system accuracy requirement is met, saving the neural network model and applying it; S8, if the radio positioning system accuracy requirement is not met, conducting the next round of training of the neural network model, returning to execute steps S4-S7 until the radio positioning system accuracy requirement is met, saving the neural network model and applying it.

[0006] As above, a drone or an aerial user is used to collect radio spectrum data in the area to be established, and the kth group of data collected during the collection process is O k =(R k ,q k ), where R k Indicates that at position q k Radio spectrum data.

[0007] As above, where R k =(rsrp k ,sinr k ,pcid k ,cellid k ), rsrp k represents the reference signal received power of the kth group of data received from the ground base station, sin r k pcid represents the signal to interference plus noise ratio of the kth group of data received from the ground base station k Indicates the physical cell identifier of the base station of the kth group of data, cellid k Indicates the logical cell address of the base station of the kth group of data.

[0008] As above, where q k =(ln g k ,lat k ) represents the location coordinates of the drone when it collects the kth set of data, where ln g kIndicates the longitude of the location where the drone collected the kth set of data, lat k Indicates the latitude of the location where the UAV was when collecting the kth set of data.

[0009] As above, the collected radio spectrum data is processed, specifically including the following sub-steps: cleaning the radio spectrum data; in response to the radio spectrum data being cleaned, normalizing the cleaned radio spectrum data; in response to the normalization of the cleaned radio spectrum data, processing and outputting the normalized radio spectrum data.

[0010] As above, the normalized radio spectrum data is processed, including: for the position coordinates q in the collected radio spectrum data k =(ln g k ,lat k ) The location coordinates after converting from longitude and latitude to meters It is expressed as: Latitude_m means converting latitude into coordinates in meters, and Longitude_m means converting longitude into coordinates in meters.

[0011] As above, after converting the location coordinates from longitude and latitude to meters, the collected and processed radio spectrum data are randomly shuffled, and 80% of the data are randomly selected as training set output, and the remaining 20% ​​of the data are used as verification set output.

[0012] As above, in the process of training radio data, the gradient descent method is used to update the model parameters of the neural network model. When the k-th batch of data is trained in the m-th round, the model parameters of the neural network model are updated as follows:

[0013]

[0014] in is the gradient of the loss function during training, where lr is the learning rate, is the model parameter value of the k-th batch of data during the m-th round of training, is the model parameter value updated after the mth round of training for the kth batch of data, and N represents the N groups of data in the neural network fed into the neural network model each time. is the radio spectrum data that needs to be trained, The location coordinates of the training radio spectrum data after being converted from longitude and latitude to meters.

[0015] As above, the neural network model performance evaluation specifically includes the following sub-steps: inputting the test set data into the neural network model trained by the training set; in response to inputting the test set data into the neural network model trained by the training set, obtaining the size of the mean square error function MSE of the test set data; performing neural network model performance evaluation according to the function value of the mean square error function MSE; if the function value of the mean square error function MSE is greater than a specified threshold, the neural network model performance evaluation result fails, and the model parameters of the neural network model are adjusted.

[0016] An aerial radio map positioning system based on deep learning specifically includes: a collection unit, a processing unit, a neural network model building unit, a training unit, an evaluation unit, a judgment unit and a storage and application unit; the collection unit is used to collect radio spectrum data; the processing unit is used to process the collected radio spectrum data; the neural network model building unit is used to build a neural network model according to the processed radio spectrum data; the training unit is used to input the processed radio spectrum data into the built neural network model to train the neural network model; the evaluation unit is used to evaluate the performance of the neural network model after completing a round of neural network model training; the judgment unit is used to judge whether the accuracy requirement of the radio positioning system is met if the result of the performance evaluation of the neural network model is passed; the storage and application unit is used to save and apply the neural network model if the accuracy requirement of the radio positioning system is met.

[0017] This application has the following beneficial effects:

[0018] (1) This application uses a deep learning algorithm and a data-based radio positioning system to avoid the limitations of traditional radio positioning systems. It no longer needs to understand the radio propagation environment, the sender and receiver of the radio signal, and is more universal.

[0019] (2) When designing the network model of the radio positioning system, the present application sets a reasonable number of neural network layers, number of neurons and hyperparameters of the neural network to accelerate the convergence speed of the algorithm and improve the accuracy of the radio positioning system.

[0020] (3) This application is directed to an intelligent radio positioning system, where users can download radio maps online or offline for positioning, pre-design of mission trajectories, and other applications, thereby saving the consumption of real resources. In addition, users who use the radio positioning system online can also share the radio spectrum data they collect, thereby increasing the data of the radio positioning system and further improving the accuracy of the radio system positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is an internal structure diagram of an aerial radio map positioning system based on deep learning provided according to an embodiment of the present application;

[0023] Figure 2 This is a flowchart of an aerial radio map positioning method based on deep learning provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are 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 those skilled in the art without creative work are within the scope of protection of the present application.

[0025] This application proposes an aerial radio map positioning method and system based on deep learning, which uses radio spectrum information to achieve accurate positioning of aerial users. This method overcomes the problem of inaccurate user positioning based on existing geometric methods, and no longer requires understanding of information such as the signal transmitted by the transmitting source and the angle between the transmitting source and the receiving source. It only needs to know information such as the radio spectrum of the receiving location.

[0026] Embodiment 1

[0027] like Figure 1 As shown, an aerial radio map positioning system based on deep learning is provided by the present application, which specifically includes: a collection unit 110, a processing unit 120, a neural network model building unit 130, a training unit 140, an evaluation unit 150, a judgment unit 160 and a storage application unit.

[0028] The collecting unit 110 is used to collect radio spectrum data.

[0029] The processing unit 120 is connected to the collecting unit 110 and is used to process the collected radio spectrum data.

[0030] The neural network model building unit 130 is connected to the processing unit 120 and is used to build a neural network model according to the processed radio spectrum data.

[0031] The training unit 140 is connected to the neural network model building unit 130 and is used to input the processed radio spectrum data into the built neural network model to train the neural network model.

[0032] The evaluation unit 150 is connected to the training unit 140 and is used to evaluate the performance of the neural network model after completing a round of training of the neural network model.

[0033] The judging unit 160 is connected to the evaluating unit 150 and is used to judge whether the accuracy requirement of the radio positioning system is met if the performance evaluation result of the neural network model is passed. If the performance evaluation result of the neural network model is not passed, the evaluating unit 150 adjusts the parameters of the neural network model.

[0034] The saving and applying unit 170 is connected to the judging unit 160 and the training unit 140 respectively, and is used to save and apply the neural network model if the accuracy requirement of the radio positioning system is met, and return to perform the next round of training of the neural network model if the accuracy requirement of the radio positioning system is not met.

[0035] Embodiment 2

[0036] like Figure 2 As shown, this application provides an aerial radio map positioning method based on deep learning, which specifically includes the following steps:

[0037] Step S210: Collect radio spectrum data.

[0038] Among them, drones or aerial users are used to collect radio spectrum data in the area where establishment is required.

[0039] Specifically, in the process of a drone or an aerial user collecting aerial radio spectrum data, it is assumed that the kth group of data collected is O k =(R k ,q k ).

[0040] Where R k =(rsrp k ,sinr k ,pcid k ,cellid k ) means at position q k Radio spectrum data, including rsrp k (Reference Signal Receiving Power) represents the reference signal receiving power of the kth group of data received from the ground base station, sin r k(Signal to Interference plus Noise Ratio) represents the signal to interference plus noise ratio of the kth group of data received from the ground base station, pcid k (Physical Cell Identifier) ​​indicates the physical cell identifier of the base station of the kth group of data, cellid k (Cell Group Identifier, base station logical cell address) represents the base station logical cell address of the kth group of data.

[0041] Among them, q k =(ln g k ,lat k ) represents the location coordinates of the drone when it collects the kth set of data, where ln g k Indicates the longitude of the location where the drone collected the kth set of data, lat k Indicates the latitude of the location where the UAV was when collecting the kth set of data.

[0042] Furthermore, the drone or aerial user collects the aerial radio spectrum data using a grid sampling method, that is, when the drone collects the radio spectrum data in the air, a set of data is collected every Δs distance during the flight.

[0043] in is the flying speed of the UAV, and Δt is the time interval between two sets of data sampling.

[0044] Through the above collection method, the granularity of the sampled data can be adjusted by adjusting the flight speed of the drone or the data sampling interval. Compared with the aimless collection of data by drones, it can reduce the cost of the sampling process.

[0045] Step S220: Process the collected radio spectrum data.

[0046] Specifically, after the drone completes collecting the radio spectrum information, in order to facilitate the next step of training, the collected data needs to be further processed. Step S220 specifically includes the following sub-steps:

[0047] Step S2201: Clean the radio spectrum data.

[0048] Radio spectrum data cleaning is to remove unreasonable radio spectrum data collected. This is mainly because drones collect ground data during takeoff and landing. If this data is used in the training of the neural network model, it will affect the accuracy of the radio positioning system in air positioning.

[0049] Step S2202: In response to the radio spectrum data being cleaned, normalizing the cleaned radio spectrum data.

[0050] After eliminating invalid radio spectrum data, the collected radio spectrum data needs to be normalized. Normalizing the radio spectrum data can help speed up the convergence of the algorithm in the subsequent training process, and thus find the optimal solution of the positioning system more quickly.

[0051] The radio spectrum data is normalized using the mean normalization method.

[0052] Specifically, for the kth group of radio spectrum data R collected k =(rsrp k ,sinr k ,pcid k ,cellid k ), its eigenvalue rsrp k 、sinr k 、pcid k 、cellid k Perform mean normalization processing respectively.

[0053] The kth group of radio spectrum data after mean normalization is: in in is the reference signal received power in the kth group of data after mean normalization, rsrp ave is the average value of all reference signal received powers collected, max{rsrp k} is the maximum value of all reference signal received powers collected. Similarly, in is the signal to interference plus noise ratio in the kth group of data after mean normalization, sin r ave is the average value of all collected signal to interference plus noise ratio, max{sin r k} is the maximum value of all collected signal to interference plus noise ratios. in is the physical cell identifier in the kth group of data after mean normalization, BSid ave is the average value of all physical cell identifiers collected, max{BSid k} is the maximum value among all collected physical cell identifiers. in is the logical cell address of the base station in the kth group of data after mean normalization, cellid aveis the average value of all collected base station logical cell addresses, max{cellid k} is the maximum value among all the collected base station logical cell addresses.

[0054] Step S2203: In response to the normalization processing of the cleaned radio spectrum data, the normalized radio spectrum data is processed and output.

[0055] The processing of the normalized radio spectrum data is to convert the unit of the normalized radio spectrum data. The radio spectrum data after the unit conversion is the final radio spectrum data output.

[0056] The object of the conversion of the radio spectrum data unit is the position coordinate q in the collected radio spectrum data. k =(ln g k ,lat k ), in order to facilitate calculation, we need to k =(ln g k ,lat k ) is converted from longitude and latitude to meters.

[0057] Specifically, assuming that the earth is approximately a sphere with a radius of R, the circumference of the earth L can be approximated as L = 2*π*R. For any latitude Lat, its circumference around the earth is Lat_l = L*cos(Lat). Since the longitude of the earth is equal to the circumference of the earth, for any longitude, its circumference is lng_l = L. Since a circle is 360°, the actual number of meters per unit degree in any latitude is Lat_C = Lat_l / 360°. Correspondingly, for any longitude, the actual number of meters per unit degree is Lng_C = Lng_l / 360°. Based on this, the actual coordinates can be calculated based on the longitude and latitude.

[0058]

[0059] Among them, Latitude_m represents the coordinate converted from latitude to meter, and Longitude_m represents the coordinate converted from longitude to meter. Based on this, the position coordinate converted from longitude and latitude to meter is expressed as: The radio spectrum data is represented as

[0060] Furthermore, after converting the location coordinates from longitude and latitude to meters, the collected and processed radio spectrum data are randomly shuffled, and then 80% of the data are randomly selected as training set output, and the remaining 20% ​​of the data are used as validation set output.

[0061] The role of the training set is to provide training data for the generation of the radio positioning system model, and the role of the validation set is to test whether the positioning of the radio positioning system model generated by the training set is accurate.

[0062] Preferably, the training set and validation set never intersect.

[0063] Step S230: Building a neural network model based on the processed radio spectrum data.

[0064] After the collected radio spectrum data is processed, a reasonable neural network model is built based on the processed radio spectrum data. The neural network model is built based on the basic neural network model.

[0065] The basic model of neural network is mainly composed of three layers, namely input layer, hidden layer and output layer. The function of the input layer is mainly to activate the input signal, the function of the hidden layer is mainly to extract features, and the linear relationship between the input layer and the output layer is represented by the neural network. The function of the output layer is mainly to output the result.

[0066] In the process of building a neural network model, the activation function of the hidden layer of the neural network of the basic neural network model in this embodiment selects the Relu activation function, which has the characteristics of fast calculation speed, rapid convergence speed, and no gradient disappearance. Since the position information needs to be accurately output in the end, the activation function of the output layer selects the liner activation function. The optimizer of the neural network selects the adam optimizer, which has the advantages of simple implementation, efficient calculation, and parameter update is not affected by the scaling transformation of the gradient.

[0067] Wherein step S230 specifically includes the following sub-steps:

[0068] Step 2301: Initialize parameters.

[0069] The main task is to initialize the model parameters of the neural network model, that is, to randomly generate the weights and biases of the neural network.

[0070] Step 2302: Set the number of hidden layers of the neural network.

[0071] The first step is to design the number of hidden layers of the neural network.

[0072] Step 2303: Set the number of neurons in the input layer, each hidden layer, and the output layer.

[0073] Since there are four characteristic values ​​of the collected radio spectrum data, the number of neurons in the input layer of the neural network is four, and since the output data is the position coordinates, the number of neurons in the output layer is two. The number of neurons in the hidden layer is randomly set.

[0074] The neurons in the hidden layer can learn the input vector With the output vector The nonlinear relationship between them can finally be obtained by inputting unlearned data To predict the corresponding position coordinates q pre .

[0075] The following is how the neural network learns the input vector through the hidden layer With the output vector The nonlinear relationship between them, where the states of neurons in the hidden layer for:

[0076]

[0077] where ν in is the weight matrix of the connection strength between the input vector and the neuron in the hidden layer, w is the weight matrix of the training data, is the bias vector, and the function σ(g) is the activation function of the neural network.

[0078] So, given the state of a neuron Then we can get the output vector in:

[0079]

[0080] where ν out is the weight matrix connecting the output vector and neurons in the hidden layer of the neural network, is its bias vector.

[0081] Step 2304: Set hyperparameters.

[0082] The hyperparameters include learning rate and loss function.

[0083] First, a small learning rate is set, and the mean absolute error (MAE) loss function is selected as the loss function.

[0084] The total loss function during training is defined as follows:

[0085]

[0086] f(w,x i ,y i ) is the loss function of the designed neural network, where w is the model parameter of the neural network, x i is the input vector of data, y i is the output vector of the neural network, N is the size of each batch of training data. The training data is part of the radio spectrum data used as training data.

[0087] It is worth noting that there is no order in which steps S2301 - S2304 are executed.

[0088] Step S2305: In response to completing the setting of the above parameters, the neural network model is built.

[0089] By setting the parameters in steps S2301-S2034, the neural network model is built.

[0090] Step S240: In response to the completion of the neural network model, the processed radio spectrum data is input into the constructed neural network model to train the neural network model.

[0091] This embodiment trains the collected data based on deep learning, and the training process aims to reduce the difference between the position predicted based on the radio spectrum data and the actual position, that is, to minimize the total loss function.

[0092] Before the processed radio spectrum data is input into the constructed neural network model, the positioning requirement accuracy γ of the radio map is set.

[0093] After setting the required positioning accuracy γ of the radio map, part of the radio spectrum data as a training set is input into the neural network model in batches for training.

[0094] Specifically, N sets of radio spectrum data are used as training sets. Input the neural network through the input layer of the neural network model, specifically input the data value of the radio spectrum and its corresponding position coordinates in is the radio spectrum data that needs to be trained, are the labels corresponding to the training radio spectrum data.

[0095] Preferably, assuming that N groups of data are fed into the neural network each time, the selection of N can be adjusted according to the effect of subsequent training, and a suitable N can speed up the training.

[0096] Specifically, during the training of radio data, the gradient descent method is used to update the model parameters of the neural network model. When the k-th batch of data is trained in the m-th round, the model parameters of the neural network model are updated as follows:

[0097]

[0098] in is the gradient of the loss function during training, where lr is the learning rate, is the model parameter value of the k-th batch of data during the m-th round of training, It is the updated model parameter value of the kth batch of data after the mth round of training. N represents N groups of data fed into the neural network each time. is the radio spectrum data that needs to be trained, are the location coordinates corresponding to the training radio spectrum data.

[0099] In response to completing a round of training of the neural network model, step S250 is executed.

[0100] Step S250: Perform neural network model performance evaluation.

[0101] After completing a round of training of radio spectrum data in the training set, the processed test set radio spectrum data is used to test the quality of the radio positioning system model trained by the training set, wherein step S250 includes the following sub-steps:

[0102] Step S2501: Input the test set data into the neural network model trained by the training set.

[0103] Specifically, the part processed in step S220 is input as the radio spectrum data of the test set into the neural network model that has completed a round of training.

[0104] Step S2502: In response to inputting the test set data into the neural network model trained by the training set, the size of the mean square error function MSE of the test set data is obtained.

[0105] The MSE function is specifically expressed as:

[0106]

[0107] Where M is the data size in the test set, is the kth * The location coordinates of the group data, is the kth predicted by the radio positioning system * The location coordinates of the group data.

[0108] Step S2503: Evaluate the performance of the neural network model based on the size of the MSE function.

[0109] The size of the verified MSE function value can indicate the quality of the trained neural network model, that is, the smaller the MSE function value verified by the verification set, that is, if the MSE function value is less than the set specified threshold, it means that the performance of the neural network model is better, the result of the neural network model performance evaluation is passed, and the positioning of the radio positioning system is more accurate, then execute step S260.

[0110] Further, according to the size of the verified MSE function value, if the MSE function value is greater than the specified threshold, it means that the performance of the neural network model is worse, and the result of the neural network model performance evaluation is failure, then step S2504 is executed.

[0111] Step S2504: Adjust the model parameters of the neural network model.

[0112] By adjusting the model parameters of the neural network model, the size of the function value of the verified MSE is gradually reduced, thereby improving the accuracy of the radio positioning system.

[0113] The model parameters for building the neural network model are specifically adjusted to adjust the parameters required for building the neural network model in steps S2301-S2304, specifically adjusting the weights and biases of the neural network, the number of hidden layers of the neural network, the input layer, the number of neurons in each hidden layer, the output layer, and the hyperparameters. After the parameter adjustment is completed, steps S240-S250 are executed again until the function value of MSE is less than the specified threshold, the adjustment of the neural network model is completed, and step S260 is executed.

[0114] Step S260: Determine whether the accuracy requirement of the radio positioning system is met.

[0115] It is determined whether the requirement MSE≤γ is met, where γ is the positioning requirement precision of the preset radio map. If the requirement is met, step S270 is executed. If the condition MSE≤γ is not met, the number of iterations is increased by 1, and the process returns to step S240 to continue the next round of training of the neural network model, that is, directly returns to execute steps S240-S250 until the mean square error function MSE≤γ of the test set data, the neural network model ends training, and step S270 is executed.

[0116] Step S270: Save the neural network model and apply it.

[0117] Apply the trained model. Users of the radio positioning system can choose to download it offline and then complete the tasks they need. In addition, users can also access the radio positioning system online to provide the radio spectrum data they collect, thereby improving the accuracy of the radio positioning system.

[0118] Wherein step S270 specifically includes the following sub-steps:

[0119] Step S2701: The user downloads the radio map positioning system from the cloud or APP.

[0120] Step S2702: The user inputs the radio spectrum data of the current location.

[0121] Step S2703: The user obtains the current location.

[0122] This application has the following beneficial effects:

[0123] (3) This application uses a deep learning algorithm and a data-trained radio positioning system to avoid the limitations of traditional radio positioning systems. It no longer needs to understand the radio propagation environment, the transmitter and receiver of the radio signal, and is more universal.

[0124] (4) When designing the network model of the radio positioning system, the present application sets a reasonable number of neural network layers, number of neurons and hyperparameters of the neural network to accelerate the convergence speed of the algorithm and improve the accuracy of the radio positioning system.

[0125] (5) This application is directed to an intelligent radio positioning system, where users can download radio maps online or offline for positioning, pre-design of mission trajectories, and other applications, thereby saving the consumption of real resources. In addition, users who use the radio positioning system online can also share the radio spectrum data they collect, thereby increasing the data of the radio positioning system and further improving the positioning accuracy of the radio system.

[0126] Although the present application is described with reference to examples, this is for illustrative purposes only and is not intended to limit the present application, and changes, additions and / or deletions to the embodiments may be made without departing from the scope of the present application.

[0127] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A deep learning-based aerial radio map positioning method, characterized in that: The specific steps include: S1, collect radio spectrum data; S2, processes the collected radio spectrum data; S3, building a neural network model based on the processed radio spectrum data; S4, in response to the completion of building the neural network model, inputting the processed radio spectrum data into the built neural network model to train the neural network model; S5, in response to completing a round of training of the neural network model, performing a performance evaluation of the neural network model; S6, if the result of the neural network model performance evaluation is passed, determine whether the accuracy requirement of the radio positioning system is met; S7, if the accuracy requirement of the radio positioning system is met, the neural network model is saved and applied; S8, if the accuracy requirement of the radio positioning system is not met, then the next round of training of the neural network model is performed, and the steps S4-S7 are returned to be executed until the accuracy requirement of the radio positioning system is met, and the neural network model is saved and applied; The radio spectrum data is collected by using drones or aerial users in the area to be established. The kth group of data collected during the collection process is Ok=(Rk,qk), where R k Indicates that at position q k Radio spectrum data; The collected radio spectrum data is processed, which specifically includes the following sub-steps: Perform radio spectrum data cleaning; In response to the radio spectrum data being cleaned, performing normalization processing on the cleaned radio spectrum data; In response to normalizing the cleaned radio spectrum data, processing and outputting the normalized radio spectrum data; The normalized radio spectrum data is processed, including: for the position coordinates q in the collected radio spectrum data k =(lng k ,lat k ) The location coordinates after converting from longitude and latitude to meters It is expressed as: Among them, Latitude_m means converting latitude into coordinates in meters, and Longitude_m means converting longitude into coordinates in meters; After converting the location coordinates from longitude and latitude to meters, the collected and processed radio spectrum data are randomly shuffled, and 80% of the data are randomly selected as the training set output, and the remaining 20% ​​of the data are used as the validation set output; In the process of training radio data, the gradient descent method is used to update the model parameters of the neural network model. When the k-th batch of data is trained in the m-th round, the model parameters of the neural network model are updated as follows: in is the gradient of the loss function during training, where lr is the learning rate, is the model parameter value of the k-th batch of data during the m-th round of training, is the model parameter value updated after the mth round of training for the kth batch of data, and N represents the N groups of data in the neural network fed into the neural network model each time. is the radio spectrum data that needs to be trained, The location coordinates of the corresponding training radio spectrum data after being converted from longitude and latitude to meters.

2. The deep learning-based aerial radio map positioning method according to claim 1, characterized in that: Where Rk=(rsrpk,sinrk,pcidk,cellidk),rsrp k represents the reference signal received power of the kth group of data received from the ground base station, sinr k pcid represents the signal to interference plus noise ratio of the kth group of data received from the ground base station k Indicates the physical cell identifier of the base station of the kth group of data, cellid k Indicates the logical cell address of the base station of the kth group of data.

3. The deep learning-based aerial radio map positioning method according to claim 1, characterized in that: q k =(lng k ,lat k ) represents the location coordinates of the drone when it collects the kth set of data, where lng k Indicates the longitude of the location where the drone collected the kth set of data, lat k Indicates the latitude of the location where the UAV was when collecting the kth set of data.

4. The deep learning-based aerial radio map positioning method according to claim 1, characterized in that: The performance evaluation of the neural network model includes the following sub-steps: Input the test set data into the neural network model trained by the training set; In response to inputting the test set data into the neural network model trained by the training set, obtaining the size of the mean square error function MSE of the test set data; The performance of the neural network model is evaluated based on the function value of the mean square error function MSE; If the function value of the mean square error function MSE is greater than the specified threshold, the performance evaluation result of the neural network model fails, and the model parameters of the neural network model are adjusted.

5. An aerial radio map positioning system based on deep learning, characterized in that: Specifically include: Collection unit, processing unit, neural network model building unit, training unit, evaluation unit, judgment unit and storage application unit; A collection unit, used for collecting radio spectrum data; A processing unit, used for processing the collected radio spectrum data; A neural network model building unit, used for building a neural network model according to the processed radio spectrum data; A training unit, used for inputting the processed radio spectrum data into the constructed neural network model to train the neural network model; An evaluation unit, used to evaluate the performance of the neural network model after completing a round of training of the neural network model; A judgment unit, used to judge whether the accuracy requirement of the radio positioning system is met if the result of the performance evaluation of the neural network model is passed; A saving and application unit, used for saving and applying the neural network model if the accuracy requirement of the radio positioning system is met; The radio spectrum data is collected by using drones or aerial users in the area to be established. The kth group of data collected during the collection process is Ok=(Rk,qk), where R k Indicates that at position q k Radio spectrum data; The collected radio spectrum data is processed, which specifically includes the following sub-steps: Perform radio spectrum data cleaning; In response to the radio spectrum data being cleaned, performing normalization processing on the cleaned radio spectrum data; In response to normalizing the cleaned radio spectrum data, processing and outputting the normalized radio spectrum data; The normalized radio spectrum data is processed, including: for the position coordinates q in the collected radio spectrum data k =(lng k ,lat k ) The location coordinates after converting from longitude and latitude to meters It is expressed as: Among them, Latitude_m means converting latitude into coordinates in meters, and Longitude_m means converting longitude into coordinates in meters; After converting the location coordinates from longitude and latitude to meters, the collected and processed radio spectrum data are randomly shuffled, and 80% of the data are randomly selected as the training set output, and the remaining 20% ​​of the data are used as the validation set output; In the process of training radio data, the gradient descent method is used to update the model parameters of the neural network model. When the k-th batch of data is trained in the m-th round, the model parameters of the neural network model are updated as follows: in is the gradient of the loss function during training, where lr is the learning rate, is the model parameter value of the k-th batch of data during the m-th round of training, is the model parameter value updated after the mth round of training for the kth batch of data, and N represents the N groups of data in the neural network fed into the neural network model each time. is the radio spectrum data that needs to be trained, The location coordinates of the corresponding training radio spectrum data after being converted from longitude and latitude to meters.

Images