UAV-assisted wireless sensor network node localization method based on deep learning

Abstract

The present invention proposes a method for positioning wireless sensor network nodes assisted by unmanned aerial vehicles based on deep learning. The application scenario of this positioning method is an outdoor area, which is specifically manifested in that the unmanned aerial vehicle moves periodically in a parallel area above the positioning area with a fixed trajectory, and broadcasts a beacon signal at a fixed time period. The ground sensor node receives the unmanned aerial vehicle beacon signal, and the node calculates the RSSI value to form an RSSI vector, calculates the RSSI similarity between nodes, and then calculates the distance between the anchor node and the unknown node, establishes a convolutional neural network positioning model, and inputs the distance between the node and the anchor node to estimate the node position coordinates. The model can overcome the influence of environmental noise to a certain extent, and the positioning performance is greatly improved compared with the existing positioning technology. The model can take advantage of the air-to-ground communication channel and the learning ability of the convolutional neural network to provide a solution to the node positioning problem, and provide theoretical support for the application research of artificial neural network technology in the field of wireless sensor networks.

Description

Technical Field

[0001] The present invention relates to the application field of deep learning artificial neural networks, and in particular to the application of artificial neural networks in wireless sensor network node positioning. Background Art

[0002] With the popularization and advancement of artificial intelligence, machine learning, deep learning and other related technologies, AI has been embedded in all aspects of people's modern life without knowing it. Advanced technical tasks such as image recognition and classification, target detection, pedestrian trajectory estimation, natural language processing, etc. are all achieved through AI. Among them, artificial neural network technology is a representative branch technology in the field of artificial intelligence. The related research on artificial neural networks accounts for a very large proportion in the entire field of artificial intelligence technology research, and its application in people's current social life is also quite extensive.

[0003] Research related to wireless sensor networks is a hot topic in the field of wireless communication research today. Network nodes in wireless sensor networks can achieve their own positioning by communicating with each other, and can then be applied to practical projects such as environmental and engineering monitoring, target tracking, etc. The existing node positioning technologies in common wireless sensor networks can be roughly divided into positioning algorithms based on ranging and non-ranging. The representative ranging methods of ranging-based positioning algorithms include TOA (Time of Arrival) based on time of arrival positioning solution, TDOA (Time Difference of Arrival) based on time difference of arrival positioning solution, AOA (Angle of Arrival) based on angle of arrival positioning solution, and RSSI (received Signal Strength Indicator) based on received signal strength indication positioning solution. The commonly used positioning frameworks include three-sided positioning method, extended Kalman filter positioning method and maximum likelihood estimation positioning method. These commonly used algorithms have preliminarily solved the node positioning problem in wireless sensor networks.

[0004] However, since the communication between ground nodes is affected by environmental noise in the actual environment, it will cause errors in the communication data between nodes, which in turn affects the accuracy of the final positioning results of the entire positioning framework. How to improve the accuracy of the positioning system has become a hot topic discussed by researchers. The existing more mature and common positioning technologies often cannot achieve very ideal results in terms of positioning accuracy, and there is still room for improvement.

[0005] Considering that the G2G communication channel between ground nodes in a wireless sensor network and the A2G communication channel between a drone and a ground node have obvious advantages in the current positioning scenario, the accuracy of the distance measurement between the drone and the ground node through the communication between the drone node and the ground node will be much higher than the communication distance measurement between the ground nodes, and a single drone mobile node only needs one GPS device for positioning. The present invention aims at the node positioning problem in a wireless sensor network. In order to achieve higher positioning accuracy based on existing research, it is proposed to adopt the form of a drone aerial anchor node. The drone flies in the air at a fixed altitude and a fixed trajectory, and the RSSI value between the drone beacon node and the ground node is obtained through the A2G channel. The RSSI values ​​between all beacon nodes and ground nodes in each flight cycle form an RSSI vector to calculate the RSSI similarity between the anchor node and the node to be located to estimate the distance between the nodes. The estimated distance is used as the input variable of the artificial neural network. The position of the ground node is obtained through the construction of a convolutional neural network model and data calculation, thereby realizing the node positioning function in the wireless sensor network. This method takes advantage of the communication advantages of the drone's A2G channel and the powerful learning ability of artificial neural networks in the field of deep learning technology, and proposes a new idea for solving the node positioning problem in wireless sensor networks. It effectively reduces the impact of environmental error noise on ranging and positioning accuracy to a certain extent, and has broad application prospects. Summary of the invention

[0006] The technical problem to be solved by the present invention is to propose a reliable UAV-assisted ground sensor node convolutional neural network positioning model for LOS (Line-of-Sight) line-of-sight environment scenarios, taking into account the presence of environmental noise in an outdoor rectangular experimental area covered by a wireless sensor network. The model uses a UAV flying in the air as an aid, sends multiple beacon nodes at regular intervals, calculates RSSI values ​​based on beacon signals received by ground nodes to be located and fixed anchor nodes, forms RSSI vectors, and calculates RSSI similarities between nodes. The distance between the anchor node and the unknown node is calculated through RSSI similarity as input, and the node position coordinates are used as output to accurately locate the position coordinates of the unknown node, thereby providing a theoretical basis for the promotion and application of artificial neural networks in the field of wireless sensor network research.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] 1. Deployment of basic conditions for positioning scenarios, determine the positioning area of ​​the target sensor node, establish a positioning reference plane coordinate system in the distribution area of ​​the wireless sensor nodes, determine the flight altitude, flight area of ​​the drone, and the path loss index and noise standard deviation of the RSSI free path propagation model in the current environmental scenario. The flight area of ​​the drone is set as a rectangular area with a side length twice the length of the ground node distribution area as the center, and the beacon node number is large enough. This setting has been proven through experiments that when the number of beacon nodes is large enough, there is a linear relationship between the RSSI similarity and distance between ground nodes. The flight altitude of the drone refers to the area of ​​the ground test area and the flight area. Select a suitable altitude to ensure that when flying at the selected altitude, the drone communication range can cover all ground sensor nodes. Deploy ground anchor nodes in a fixed position in an n*n matrix distribution, and determine the spacing between adjacent anchor nodes as d. width , record the position coordinates of each anchor node. When the drone flies in the flight area, the distribution of ground fixed anchor nodes is:

[0009]

[0010] x i+1 -x i =y j+1 -y j =d width

[0011] 2. After determining the flight altitude and flight area of ​​the drone, select the appropriate flight trajectory of the drone. The principles for selecting the flight trajectory of the drone are:

[0012] a. The flight trajectory covers as much area as possible within the flight area, ensuring that the location of the aerial beacon nodes can be evenly and dispersedly distributed over most of the flight area;

[0013] b. The flight trajectory is set to a closed curve shape to ensure that the next flight cycle can start immediately after the end of each flight, providing convenience for the system to measure distance multiple times;

[0014] c. The beacon nodes are evenly spaced on the set UAV flight trajectory, so that the UAV can send beacon signals to the ground nodes at equal time intervals when flying at a constant speed;

[0015] After determining the flight trajectory of the drone, the distance estimation DE (Distance Estimation) module is established. The beacon signal is sent at the drone's aerial beacon position. The ground node receives the signal sent by the aerial beacon node, calculates the RSSI value, and forms an RSSI vector to return to the drone. The drone calculates the RSSI similarity between nodes based on the RSSI vector returned by the ground node, and calculates the distance between nodes based on the RSSI similarity. The specific process is expressed as follows:

[0016] (1) UAV air beacon signal sending time interval and trajectory equation:

[0017]

[0018] t transmission ={t = (k-1) * t T |k=1,2,...,K}

[0019] (2) The RSSI value of the beacon signal broadcast by drone aerial beacon node k received by ground node i is expressed as

[0020]

[0021] (3) The RSSI vector formed by the node with ground index i receiving the beacon signal sent by the drone during the mth flight cycle is

[0022]

[0023] (4) The similarity calculation and distance estimation process of the RSSI vectors collected for each pair of ground nodes is expressed as:

[0024]

[0025]

[0026] Among them, P(t) represents the coordinates of the position reached by the drone at time t, which is divided into the horizontal coordinate x(t) and the vertical coordinate y(t). T is the time interval between each beacon sent by the UAV to the ground node, t transmission is the time set when the drone air anchor node sends the beacon signal to the ground node, t k is the time node when the UAV sends a signal to the ground for the kth time, P[d 0 ] indicates that the communication distance is the reference distance d 0 The RSSI value at that time, γ represents the free path loss index in the positioning scenario environment, c k represents the kth drone beacon node location, n i represents the position of the ground node numbered i, X σ is the environmental noise random variable with standard deviation σ, The RSSI vector composed of the RSSI values ​​of the beacon signals sent by all drones received by the ground node numbered i during the mth flight of the drone is calculated. It represents the RSSI value of the beacon signal sent by the UAV when it passes the jth beacon node during the mth flight and received by the i-th ground node. It represents the RSSI similarity between the i-th node and the j-th node on the ground. K is the total number of drone aerial beacon nodes. After mathematical calculation, it is found that the RSSI similarity is the distance d between the ground nodes. ij There is a linear function relationship, where Z is the environmental noise variable, V is the area of ​​the drone's flight area, and the linear relationship between RSSI similarity and distance is transformed into d ij and A and B are constant quantities calculated based on environment and scene parameters.

[0027] Finally, the calculated distance vector between the fixed anchor node and the node to be located is rearranged to form a convolution block of a multi-channel feature map, which is input into the subsequent convolutional neural positioning network. The specific number of convolution block channels M corresponds to the number of cycles of the drone flying along a fixed trajectory in the air. The feature map in each channel numbered m is arranged in the form of an n*n matrix. Each neuron geometrically corresponds to the ground fixed anchor node in the experimental area whose arrangement position is the same as the relative position of the neuron in the feature map. The value in the neuron is the estimated straight-line distance between the ground node to be located and the ground fixed anchor node represented by the neuron. The anchor node and the corresponding number j are represented as A. j , the node to be located and the corresponding number are represented as U i ,The process of UAV’s ranging between ground nodes and data rearrangement input into the positioning system can be specifically expressed as:

[0028]

[0029] f(x)=Ax+B

[0030]

[0031]

[0032] in Represents the feature map matrix corresponding to the mth channel of the convolutional neural network input convolution block, is the estimated distance between the jth ground anchor node and the ith unknown node obtained during the test of the UAV’s mth flight. f(x) represents the distance estimation equation with the independent variable being RSSI similarity in the positioning scenario. Input iRepresents the network input when the convolutional neural network estimates the position of the i-th node on the ground, where M means that the input convolution block has M channels, that is, the drone will fly M cycles along the set trajectory, and will send a beacon signal to the ground node once when passing each aerial beacon node in each flight cycle. The RSSI values ​​calculated from each K beacon signals form an RSSI vector, and finally each ground node receives M RSSI vectors from the aerial anchor node and returns them to the drone. Each time the distance is estimated, the RSSI vector obtained in the same flight cycle is used to calculate the similarity; the internal structure of the convolutional neural network model includes an input layer, several convolutional layers, several fully connected layers and an output layer, where the number of convolutional layers and fully connected layers in the network depends on the number of anchor nodes N=n*n deployed in the actual experiment and the set convolution kernel size. The internal structure of the network and the setting of parameters will affect the running time and space complexity of the entire system and the accuracy of the system for node positioning.

[0033] 3. Build a convolutional neural network positioning network model and train and optimize it. Compared with ordinary fully connected neural networks, artificial intelligence deep learning convolutional neural network technology greatly reduces the amount of data calculation and has a strong learning ability for data features. Therefore, convolutional neural networks are used to learn the node distribution characteristics in wireless sensor networks. The obtained network model can estimate the positions of unknown nodes distributed in the experimental area, and in the case of environmental noise errors in RSSI communication ranging, it can overcome the impact of errors on positioning to a certain extent.

[0034] The size of the network's input convolution block and the value of each neuron have been described in detail in Technical Solution 2. The size of the convolution kernel, the number of hidden layers, and the selection of the activation function have the following principles:

[0035] (1) The number of drone aerial anchor nodes is directly related to the computational complexity of the convolutional neural network and the complexity of the network. Generally, n is selected to be around 3-4, that is, 9 or 16 anchor nodes are evenly deployed in the experimental area. When the drone flies over the beacon node, it sends beacon signals to these ground anchor nodes and all nodes to be located.

[0036] (2) The number of times the UAV flies along a fixed trajectory is selected to be 24-48 times, that is, the number of channels of the input layer convolution block of the positioning convolutional neural network is selected to be 24-48. The number of flight cycles means the number of times the distance between nodes is estimated. Too few ranging times will lead to large errors in network positioning. The errors of a few ranging data are accidental, which has a significant impact on positioning accuracy. Too many received signals will make the system model complex, and the accuracy improvement will not be obvious.

[0037] (3) The size of the convolution kernel is generally 2*2, with no padding and no holes, and a step size of 1. In the positioning convolutional neural network structure, the feature map of the input layer can be understood as the distance matrix between the anchor node and the node to be determined. Each element value in the matrix is ​​very important. For high-density data, a step size of 1, no holes, and no padding is selected to allow the network to learn every detail feature in the input data in detail. The convolution layer activation function selects the Relu function, and the fully connected layer uses the Sigmoid function for activation;

[0038] (4) The mean square error function is used as the loss function for training to estimate the difference between the predicted node position of the network and the actual node position, which can be specifically expressed as:

[0039]

[0040] Where N represents the number of nodes in a single Batch in the training data set. Indicates that the convolutional neural network positioning system estimates the position coordinates of the i-th node, Indicates the actual position coordinates of the i-th node in the training data set. For the N value setting problem, in order to prevent the network parameters from falling into a local optimal situation during the training process, the N value should not be too large, generally at the level of 50-100. The total number of optimizations performed in each iteration of training is the quotient obtained by dividing the total number of training samples, Samples, by N. In each iteration of training, Samples / N back-propagation operations are required.

[0041] (5) Commonly used training optimizers for convolutional neural networks include SGD, Adagrad, Adam and other built-in optimizers in deep learning frameworks. Among them, the Adam optimizer has obvious advantages over other optimizers in the application scenario of convolutional neural networks for wireless sensor network node positioning. Specifically, it uses the first-order moment estimation and second-order moment estimation of the gradient to dynamically adjust the parameter learning rate. The parameters are relatively stable, the memory requirement is small, and the update step size can be automatically maintained at a level around the initial learning rate. The step size annealing of the parameters can be naturally achieved, which is suitable for large-scale parameters and scenarios with gradient noise. Using the Adam optimizer to train the convolutional neural network for wireless sensor node positioning will obtain a network model with a relatively excellent error level, and its noise resistance and positioning performance can reach a relatively ideal state.

[0042] 4. Train the built convolutional neural network positioning system. After setting the parameters and size of each layer of the network, randomly generate a large number of actual coordinates of the nodes to be located. The software tests the RSSI vector of each node obtained according to the fixed trajectory of the drone. The RSSI similarity between nodes is calculated to estimate the test distance between the anchor node and the node to be located. A large amount of training data is obtained by the ranging module. The estimated distance is grouped and input into the network. The Batch Size of each group of input network is set. For each group of training samples input, a back propagation optimization operation is performed on the network. In each iteration, the back propagation optimization operation of the total number of samples divided by the number of samples in each batch is performed. The number of iterations is set to 800-1000, and the loss function is set to the mean square error function. As the loss function decreases, the initial learning rate is modified to decrease exponentially. Finally, the root mean square error of network positioning can reach the ideal high-precision level.

[0043] 5. Test the designed positioning system in an actual scenario. The UAV flies along a fixed trajectory in a set area at a fixed altitude above the experimental area. The ground node collects the beacon signal sent by the UAV beacon node, calculates the RSSI value to form an RSSI vector and sends it back to the UAV. The UAV calculates the RSSI similarity based on the RSSI vectors returned by different nodes and then calculates the distance between the nodes. The estimated distance between the nodes is rearranged into a convolution block and input into the trained convolutional neural positioning network. The network outputs the estimated position coordinates of the final node.

[0044] Beneficial effect: The model proposed in the present invention is a convolutional neural network model for the static node positioning scenario of an unmanned aerial vehicle (UAV)-assisted outdoor wireless sensor network. The model cleverly utilizes the powerful learning ability of the convolutional neural network technology in deep learning for large-scale data, rearranges the distance estimation data between nodes into a natural image form and inputs it into the convolutional neural network, and uses the convolutional neural network to learn the data relationship characteristics of the distance between nodes and the node position coordinates. At the same time, the communication channel performance between the UAV and the ground node is better than the mutual communication between the ground nodes, which overcomes the error effect of environmental noise on ranging to a certain extent. In the ranging process, the method of estimating the distance using the relationship between RSSI similarity and the distance between nodes replaces the free path loss model, thereby improving the ranging accuracy between nodes. On this basis, the convolutional neural network multi-channel model can overcome the influence of the existing ranging noise error on coordinate estimation to a certain extent by learning the commonalities between multiple ranging data. In the actual wireless sensor network node positioning scenario, the model can obtain the RSSI information of the communication between the drone anchor node and the ground node through a reliable and efficient air-to-ground channel, and process the information into a convolution block as the network input. The convolutional neural network positioning model calculates the estimated node position coordinates and can obtain a relatively accurate position estimation output, which provides theoretical support for researchers to design and study wireless sensor network node positioning problems and promote the application of deep learning artificial neural network models in wireless sensor network communications. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a scenario structure diagram of the deep learning algorithm for drone-assisted wireless sensor network node localization;

[0046] Figure 2 It is the internal flow chart of DE module work;

[0047] Figure 3 It is a diagram of the internal structure of the localization convolutional neural network;

[0048] Figure 4 It is a flowchart of the establishment and training process of the positioning model in the offline stage;

[0049] Figure 5 It is a process flow chart of the proposed positioning algorithm performing positioning work in the online stage. DETAILED DESCRIPTION

[0050] 1. Positioning scenario

[0051] The scene model used by the positioning network proposed in the present invention is as follows: Figure 1 As shown:

[0052] (1) In a square experimental area covered by a wireless sensor network, static sensor nodes and fixed anchor nodes are randomly deployed on the ground. A drone flies in a parallel plane area with an area four times the size of the experimental area. During the flight, it periodically sends beacon signals to the ground nodes to ensure that the drone's communication range can cover all sensor nodes in the experimental area.

[0053] (2) Select the optimal flight altitude of the UAV, confirm the UAV flight trajectory, the number of beacon nodes, and the beacon signal sending time. Here, 1,600 beacon nodes in the air and a regular octagonal UAV flight trajectory are used. When the UAV passes each side of the regular octagon, 200 beacon signals are sent to the ground.

[0054] (3) Determine the position coordinates of the ground fixed anchor nodes, the UAV flight trajectory equation, and the number of times the UAV flies along the fixed trajectory, that is, the number of channels (depth) of the input layer of the convolutional neural network. Here, 9 anchor nodes are evenly distributed in the experimental area in a 3*3 direction in the horizontal and vertical directions. Here, it is assumed that the experimental area is a square area of ​​120m*120m. The position coordinates of the anchor nodes are:

[0055]

[0056] x i+1 -x i =y j+1 -y j =40

[0057] The flight area of ​​the drone is x, y∈(-60, 180), and the flight trajectory is set to a regular octagon. The coordinates of the 8 vertices of the regular octagon are: The trajectory of a single flight cycle can be expressed as 1 Start flying and eventually return to P 1 The process is denoted as P 1 →P 2 →P 3 →P 4 →P 5 →P 6 →P 7 →P 8 →P 1 , determine the RSSI path loss index γ of the air-to-ground channel at the UAV flight altitude, and the noise standard deviation σ in the current environmental positioning scenario, and determine to use the convolutional neural network positioning model with a channel number of M = 24.

[0058] 2. The internal working principle of the DE (Distance Estimation) module is as follows Figure 2 As shown:

[0059] The specific workflow of the DE module in the model proposed in this invention can be divided into the following steps:

[0060] (1) The UAV flies along the set trajectory over the experimental area for a total of M = 24 cycles. During the mth flight, when passing through the aerial beacon node k, it sends an RSSI beacon signal to the ground node. The RSSI value of the beacon signal received by the ground node numbered i is expressed as:

[0061]

[0062] (2) The RSSI vector corresponding to node i, which is composed of the RSSI values ​​calculated by the beacon signals collected by all beacon nodes in the mth flight cycle, is expressed as:

[0063]

[0064] (3) The DE module calculates the RSSI similarity between the ground fixed anchor node j and the unknown node i based on the input RSSI vector and then converts it into the estimated distance between the nodes:

[0065]

[0066]

[0067] in represents the RSSI similarity from the jth ground anchor node to the ith unknown node during the mth flight of the drone, represents the estimated straight-line distance between nodes during the mth drone flight, K = 1600 is the total number of drone air beacon nodes, A and B are the linear relationship constant coefficient and bias constant between RSSI similarity and distance determined by environmental factors in the current scenario;

[0068] (3) The process of estimating the distance between the ground fixed anchor node and the unknown node in multiple UAV-assisted tests is expressed as:

[0069]

[0070] in is the estimated distance between the jth ground fixed anchor node and the ith unknown node during the mth flight of the UAV, is the vector composed of the estimated distances between the ith ground node and all ground anchor nodes in the mth UAV flight. The UAV flies 24 times at the set altitude and trajectory, and obtains 24 distance vector;

[0071] (4) The calculated distance vectors between the anchor nodes and the unknown nodes are rearranged to form a convolution block of a multi-channel feature map. The specific number of convolution block channels corresponds to the number of times the drone flies along a fixed trajectory in the air. The feature map in each channel is arranged in a 3*3 matrix. Each neuron in the feature map corresponds to the drone air anchor node whose arrangement position in the experimental area is the same as the relative position of the neuron in the feature map. The value in the neuron is the straight-line distance between the ground node to be located and the corresponding ground anchor node and the node to be located, which can be expressed as:

[0072]

[0073] f(x)=Ax+B

[0074]

[0075]

[0076] in Represents the feature map matrix corresponding to the mth channel of the convolutional neural network input convolution block, which is the estimated distance obtained by collecting RSSI vector values ​​and calculating RSSI similarity Rearranged, input i represents the network input of the convolutional neural network when estimating the position of the i-th node. Multiple feature maps obtained by multiple flights of the drone are superimposed into multiple channels to form a convolution block of the input positioning network. The input convolution block has 24 channels, that is, the drone will fly along the set trajectory for 24 cycles. In each flight cycle, it will send a beacon signal to the ground node when passing each beacon node. K = 1600 beacons are sent in each flight cycle. Finally, the ground node receives a total of 24 signals from the aerial anchor node during all flights. vector, where m = 1, 2, ..., 24, i = 1, 2, ..., N, N is the total number of ground nodes to be located, and the output convolution block visualization graph of the DE module is Figure 2 As shown below, this convolutional block is used as input to the convolutional neural localization network.

[0077] 3. The internal structure of the convolutional neural localization network is as follows Figure 3 As shown:

[0078] The network consists of an input layer, 4 hidden layers and an output layer. The first two hidden layers are convolutional layers, and the last two hidden layers are fully connected layers. After the convolutional neural positioning network model is built according to the scenario requirements, appropriate network parameters, learning rate, optimizer and loss function are set. A large number of randomly generated training data samples are used to perform multiple iterative training on the network. Finally, a positioning network with excellent performance is obtained and applied to actual positioning scenarios to estimate the position of unknown nodes.

[0079] 4. The establishment and training process of the proposed convolutional neural localization network model based on deep learning is as follows: Figure 4 As shown, it is divided into the following steps:

[0080] (1) After determining the positioning environment and scenario, randomly generate a large amount of training data and test data that match the scenario;

[0081] (2) According to the determined scenario, i.e., the UAV flight altitude, trajectory, number of beacons, and the location coordinates of the ground anchor nodes, a convolutional neural network positioning model that meets the scenario is built;

[0082] (3) Use a large number of random training samples to train the established network model;

[0083] (4) After multiple iterations, the network positioning error meets the positioning requirements and a trained model is obtained;

[0084] 5. The whole system workflow of the proposed UAV-assisted wireless sensor node positioning deep learning algorithm is shown in the figure below. Figure 5 As shown, it is divided into the following steps:

[0085] (1) The drone flies along a fixed altitude and trajectory and periodically broadcasts beacon signals to ground nodes;

[0086] (2) The ground node receives the beacon signal broadcast by the drone, calculates the RSSI value to form an RSSI vector, and inputs it into the DE module for data processing;

[0087] (3) The DE module inputs the processed data into the trained convolutional neural localization network model;

[0088] (4) Convolutional neural network estimates the position coordinates of unknown nodes.

Claims

1. A UAV-assisted wireless sensor network node localization method based on deep learning, It is characterized in that The steps include: Step 1: Deployment of basic conditions for the positioning scenario, determine the positioning area of ​​the target sensor node, establish a positioning reference plane coordinate system in the distribution area of ​​the wireless sensor nodes, determine the flight altitude and flight area of ​​the drone, and the path loss index and noise standard deviation of the RSSI free path propagation model in the current environment scenario, deploy ground anchor nodes in a fixed position in an n*n matrix distribution, and record the position coordinates of each anchor node to ensure that the communication range of the drone can cover all ground sensor nodes when flying in the flight area; Step 2: Set the fixed flight trajectory and speed of the drone, the signal transmission cycle and the number of beacons. According to the experimental simulation test, the drone flies along the set trajectory and sends beacon signals to the ground nodes at the specified position according to the set signal transmission cycle. The ground nodes receive the beacon signals and calculate the RSSI values ​​to form RSSI vectors. The linear relationship between the RSSI vector calculation similarity between the ground nodes and the distance between the nodes is obtained to obtain the RSSI similarity-distance estimation curve. Step 3: Establish a convolutional neural network model, where the height and width of each feature map of the input variable are n, the number of neurons is n*n, the input value of each neuron is the estimated distance between the ground node to be located and the corresponding anchor node, the number of channels is the number of times the drone distance estimation module performs distance estimation, and through two convolutional layers and two fully connected layers, the position coordinates of the node to be located are finally output; Step 4: Establish training set data to train the convolutional neural network model, select appropriate optimizers, learning rates and iteration times, update network parameters to improve the positioning performance of the network, continuously reduce the error between the network's position estimation coordinates and the true coordinates of the nodes, and perform performance evaluation and error analysis on the obtained convolutional neural network model; Step 5: Collect the beacon signals received by the ground anchor node and the drone to be located, calculate the RSSI value to obtain the RSSI vector, calculate the RSSI similarity between the anchor node and the node to be located, estimate the straight-line distance between the anchor node and the node to be located through the RSSI similarity-distance curve, process and arrange them into a convolution feature block pattern, and input the trained convolution neural network model to obtain the position coordinate estimation of the ground node in the wireless sensor network; In step 1, the UAV flight altitude is selected as the flight altitude with the minimum transmission path loss index of the air-to-ground channel. The UAV flight area is set to be a two-dimensional plane in the air that is parallel to the distribution area of ​​the ground sensor nodes and the area of ​​the area is 4 times the ground experimental area. The distribution of the ground fixed anchor nodes is set to be n*n horizontal and vertical axes are evenly distributed in the wireless sensor network coverage experimental area, and the spacing between adjacent anchor nodes is equal, which is represented by d width , each fixed anchor node is numbered A p , where p = 1, 2, ..., n*n, and the ground fixed anchor node deployment coordinate distribution is: x i+1 -x i =y j+1 -y j =d width In step 2, the fixed flight trajectory of the UAV is set as a closed curve, that is, after one cycle of flight, it returns to the flight starting point for the next cycle of flight. The total number of beacon nodes is set to K, and K beacon nodes are evenly and equidistantly distributed on the UAV flight trajectory P(t). The signal sending period is t T Set as the time interval from the drone flying from one beacon to the next adjacent beacon. The ground node i receives all beacon signals in the air during the mth flight and obtains the RSSI value to form the RSSI vector RSSI i m , the RSSI similarity between nodes is calculated based on the RSSI vector, and then the distance between nodes is calculated, which is specifically expressed as: (1) The trajectory equation of a UAV flying in the air: t transmission ={t=(k-1)*t T |k=1,2,...,K} (2) The RSSI value obtained by ground node i receiving the beacon signal broadcast by drone aerial beacon node k is expressed as (3) The RSSI vector formed by the node with ground index i receiving the beacon signal sent by the drone during the mth flight cycle is (4) For each pair of ground anchor node j and unknown node i, the RSSI vectors are collected and the similarity calculation and distance estimation process is expressed as: Among them, P(t) represents the coordinates of the position reached by the drone at time t, which is divided into the horizontal coordinate x(t) and the vertical coordinate y(t). T is the time interval between each beacon sent by the UAV to the ground node, t transmission is the time set when the drone air anchor node sends the beacon signal to the ground node, t k is the time node when the UAV sends a signal to the ground for the kth time, P[d 0 ] indicates that the communication distance is the reference distance d 0 The RSSI value between the UAV and the ground node at that time, γ represents the free path loss index in the positioning scenario environment, c k represents the kth drone beacon node location, n i represents the position of the ground node numbered i, X σ is the environmental noise random variable with standard deviation σ, It represents the RSSI vector composed of the RSSI values ​​of beacon signals sent by all drones received by the ground node numbered i during the mth flight of the drone. It represents the RSSI value of the beacon signal sent by the UAV when it passes the jth beacon position during the mth flight and received by the i-th ground node. It represents the RSSI similarity between the i-th node and the j-th node on the ground. K is the total number of drone aerial beacon nodes. After mathematical calculation, it is found that the RSSI similarity is the distance d between the ground nodes. ij There is a linear function relationship, where Z is the environmental noise variable, V is the area of ​​the drone's flight area, and the linear relationship between RSSI similarity and distance is transformed into d ij and A and B are constant quantities, calculated based on environment and scene parameters; In step 3, the structure of the input convolutional block of the constructed convolutional neural network positioning model is expressed as: in Represents the feature map matrix corresponding to the mth channel of the convolutional neural network input convolution block, is the estimated distance between the jth ground anchor node and the ith unknown node obtained during the test of the UAV’s mth flight. i Represents the network input when the convolutional neural network estimates the position of the i-th node on the ground, where M represents that the input convolution block has M channels, that is, the drone will fly M cycles along the set trajectory, and will send a beacon signal to all ground nodes once when passing each aerial beacon position in each flight cycle, and the ground nodes receive the beacon signal sent by the drone to calculate the RSSI value. Finally, each ground node receives M RSSI vectors from the aerial anchor node and returns them to the drone. Each distance estimation uses the RSSI vector obtained in the same flight cycle to calculate the similarity; the internal structure of the convolutional neural network model includes an input layer, several convolutional layers, several fully connected layers and an output layer, where the number of convolutional layers and fully connected layers in the network depends on the number of anchor nodes N=n*n deployed in the actual experiment and the set convolution kernel size. The internal structure of the network and the setting of parameters will affect the running time and space complexity of the entire system and the accuracy of the system for node positioning; In step 4, during the training process of the convolutional neural network positioning system, the training set data uses computer randomly generated data, the RSSI value of the beacon signal received by the node is estimated using the free path loss model and the node position, the RSSI vector is generated, and the RSSI similarity between nodes is calculated to estimate the distance to train the network. The selection of different optimizers and network parameters greatly affects the final positioning performance of the network. Among all the optimizers included in the pytorch deep learning framework, the selection of the Adam optimizer for network training has obvious advantages in the huge amount of node positioning data and actual positioning scenarios; the mean square error function is selected as the loss function for network training: Where N represents the number of nodes in a single Batch in the training data set. Indicates that the convolutional neural network positioning system estimates the position coordinates of the i-th node, Represents the actual position coordinates of the i-th node in the training data set; In step 5, the UAV flies at a fixed height and in a fixed area above the experimental area according to the set trajectory, and sends beacon signals periodically. The ground node collects the RSSI beacon signals sent by the UAV to form an RSSI vector and returns it to the UAV. The UAV calculates the distance between the anchor node and the node to be located based on the RSSI vectors returned by different nodes, and after rearranging it into a feature map superposition to form a multi-channel convolution block, it is input into the trained positioning network to estimate the position coordinates of the ground node to be located.

Images