Unmanned aerial vehicle base station RSSI positioning method and system based on deep learning
By using a single UAV carrying a base station and combining it with deep learning algorithms, the high cost and low efficiency of traditional multi-base station positioning methods are solved, achieving high-precision terminal positioning in complex environments with real-time performance and scalability.
Patent Information
- Application Number
- CN202510964844.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-10-31
AI Technical Summary
Traditional terminal positioning methods require multiple base stations, which are costly, inefficient, and have unstable positioning accuracy in complex environments, making them difficult to adapt to complex terrain and obstacles.
By using a single UAV carrying a base station and combining it with deep learning algorithms, a hybrid deep learning model is designed to achieve high-precision positioning of the terminal by collecting the terminal's RSSI signal strength and the UAV's GPS coordinates. An edge-cloud collaborative learning system is then used for cross-scenario adaptation.
It achieves high-precision, low-cost terminal positioning, adapts to complex environments, and possesses real-time and scalability capabilities, thereby reducing system complexity and deployment costs.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicle (UAV) application technology, and in particular to a method and system for UAV base station RSSI (Received Signal Strength Indication) positioning based on deep learning. This method enables high-precision satellite-free positioning of UAVs in complex scenarios and is applicable to scenarios such as indoor inspection, disaster relief, and urban logistics. It is also suitable for the precise positioning of target terminals. Background Technology
[0002] With the rapid development of drone technology, drones are increasingly being used in search and rescue, monitoring, and surveying. Many applications require precise positioning of specific ground-based terminal devices, such as locating signal transmitters carried by people in distress during search and rescue operations, or locating animals carrying signal transmitters in wildlife research.
[0003] Traditional terminal positioning methods primarily rely on the triangulation principle, requiring multiple base stations to simultaneously receive terminal signals and calculate the terminal's location using parameters such as signal strength and time difference of arrival. This method suffers from the following problems in practical applications:
[0004] 1. Requires the deployment of multiple base stations, resulting in high costs and low efficiency.
[0005] 2. In environments with complex terrain or numerous obstacles, signal propagation models struggle to accurately describe [the situation].
[0006] 3. Positioning accuracy is greatly affected by environmental factors and has poor stability.
[0007] Therefore, there is an urgent need for a method that can achieve high-precision positioning using a single mobile base station. Summary of the Invention
[0008] This invention provides a method and system for RSSI positioning of unmanned aerial vehicles (UAVs) based on deep learning. By using a single UAV carrying a base station, the method collects the RSSI signal strength of the terminal at different locations and the GPS coordinates of the UAV, and combines the deep learning algorithm to achieve accurate positioning of the terminal.
[0009] 1. Technical Solution
[0010] The technical solution of the present invention includes the following steps:
[0011] 1.1 Data Acquisition Phase:
[0012] (1) The drone carrying the base station flies along a preset flight path.
[0013] (2) During flight, record the received terminal RSSI value and the corresponding UAV GPS coordinates.
[0014] (3) Collect a sufficient number of data points to form an RSSI-location dataset.
[0015] 1.2 Data Preprocessing Stage:
[0016] (1) Filter the acquired RSSI data to eliminate noise.
[0017] (2) Standardize GPS coordinates
[0018] (3) Construct time-series features and extract RSSI change trends.
[0019] 1.3 Deep Learning Model Training Phase:
[0020] (1) Design a deep neural network model suitable for RSSI localization
[0021] (2) Train the model using the collected dataset.
[0022] (3) Optimize model parameters to improve positioning accuracy.
[0023] 1.4 Terminal positioning stage:
[0024] (1) The UAV flies along a specific path and collects the terminal RSSI value in real time.
[0025] (2) Input the collected data into the trained deep learning model
[0026] (3) Predicted position coordinates of the output terminal
[0027] (4) Improve positioning accuracy by fusing multiple prediction results.
[0028] 2. Innovation Points
[0029] (1) Single UAV dynamic scanning technology
[0030] Unlike the multi-base station solution, this method uses a single drone to carry a base station and perform spiral trajectory scanning to achieve mobile data collection.
[0031] (2) Edge-Cloud Collaborative Learning System
[0032] Based on the offline model, an edge incremental learning and cloud transfer learning mechanism is introduced to support rapid adaptation of the model across different scenarios.
[0033] 3. Technical Effects
[0034] (3) Single base station positioning: Only one drone carrying a base station is needed to complete the positioning task, which greatly reduces system complexity and deployment costs.
[0035] (4) High-precision positioning: The complex relationship between RSSI and position is automatically learned through a deep learning model, overcoming the problem of inaccurate signal propagation models in traditional methods.
[0036] (5) Strong environmental adaptability: Deep learning models can adapt to the signal propagation characteristics in different environments and maintain high positioning accuracy even in complex terrain and obstacle environments.
[0037] (6) Good real-time performance: After the model is trained, the localization process requires little computation and can achieve real-time localization.
[0038] (7) High scalability: The system architecture supports multiple deep learning models and can be optimized according to specific application scenarios. Attached Figure Description
[0039] Figure 1 This is a system architecture diagram as described in the embodiments of the present invention;
[0040] Figure 2 This is a diagram of the deep learning model architecture described in the embodiments of the present invention;
[0041] Figure 3 This is a flowchart illustrating the positioning process described in this embodiment of the invention;
[0042] Figure 4 This is a system workflow diagram as described in the embodiments of the present invention. Detailed Implementation
[0043] 1. System Architecture
[0044] The system architecture of this invention mainly includes the following components:
[0045] (1) Unmanned aerial vehicle (UAV) platform: equipped with base station equipment, capable of flying along a preset path and recording its own GPS location.
[0046] (2) Base station equipment: installed on the drone, used to receive terminal signals and measure RSSI values.
[0047] (3) Terminal equipment: transmits wireless signals and is the target object for positioning.
[0048] (4) Data processing unit: can be installed on the UAV or ground station, responsible for data preprocessing and model inference.
[0049] (5) Model training unit: Usually deployed at ground stations, responsible for training and optimizing deep learning models.
[0050] (6) Communication module: Enables data transmission between the UAV and the ground station.
[0051] 2. Data Acquisition Methods
[0052] (1) Flight path design:
[0053] ① Grid-like path: The drone flies along a preset grid-like path to cover the target area.
[0054] ② Spiral path: Flying in a spiral from the outside in, suitable for gradually narrowing down the search area for the target region.
[0055] ③ Adaptive Path: Dynamically adjusts the flight path based on real-time RSSI values to improve acquisition efficiency.
[0056] (2) Data Recording:
[0057] ① Recorded items: RSSI value, UAV GPS coordinates, timestamp, flight altitude, flight speed, etc.
[0058] ② Sampling frequency: Determined based on the UAV's flight speed and required positioning accuracy, typically 1-10Hz.
[0059] ③ Data storage: Combining local storage with real-time transmission ensures data security.
[0060] (3) Calibration method:
[0061] ① Conduct flight tests in an environment with a known terminal location and collect calibration data.
[0062] ②Analyze the signal propagation characteristics under different environmental and weather conditions.
[0063] ③ Establish a basic signal propagation model as prior knowledge for deep learning.
[0064] 3. Data Processing Methods
[0065] (1) Noise filtering:
[0066] ①Median filtering: Removes burst noise from RSSI signals
[0067] ② Kalman filtering: smooths RSSI signal fluctuations
[0068] ③ Wavelet transform: Separating different frequency components in a signal
[0069] (2) Feature extraction:
[0070] ① Time-domain characteristics: Statistical features of RSSI such as mean, variance, maximum, and minimum.
[0071] ② Frequency domain characteristics: Extract the spectral characteristics of the RSSI signal through Fourier transform.
[0072] ③ Spatial characteristics: Correlation characteristics between UAV location and RSSI value
[0073] ④ Gradient characteristics: Gradient information of RSSI value as a function of location
[0074] (3) Data augmentation:
[0075] ① Add Gaussian noise: simulate signal fluctuations in a real-world environment
[0076] ② Random discard: Analog signal loss situation
[0077] ③ Synthetic data: Generating additional training data based on the signal propagation model
[0078] 4. Deep Learning Model Design
[0079] This invention employs a hybrid deep learning model, combining CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network):
[0080] (1) Model input:
[0081] ①RSSI sequence data: RSSI measurements including the time dimension
[0082] ②Location sequence data: corresponding UAV GPS coordinates
[0083] ③ Auxiliary features: Flight altitude, speed, and other related information
[0084] (2) Model structure:
[0085] ①CNN module: Extracts spatial features from RSSI-location data
[0086] - Input layer: Receives the preprocessed feature matrix
[0087] - Convolutional Layer 1: 32 3×3 convolutional kernels, ReLU activation function
[0088] - Pooling layer 1: 2×2 max pooling
[0089] - Convolutional Layer 2: 64 3×3 convolutional kernels, ReLU activation function
[0090] - Pooling layer 2: 2×2 max pooling
[0091] ②LSTM module: Captures the temporal characteristics of RSSI variations over time and location.
[0092] Input: Feature sequence extracted by CNN
[0093] -LSTM layer 1: 128 hidden units
[0094] -LSTM layer 2: 64 hidden units
[0095] ③ Fully connected layer: fuses features and maps them to the terminal location.
[0096] - Fully connected layer 1: 256 neurons, ReLU activation function
[0097] - Dropout layer: Dropout rate of 0.3 to prevent overfitting.
[0098] - Fully connected layer 2: 128 neurons, ReLU activation function
[0099] - Fully connected layer 3: 64 neurons, ReLU activation function
[0100] - Output layer: 2 or 3 neurons (depending on 2D or 3D localization), linear activation function
[0101] (3) Loss function design:
[0102] ① Main loss function: Mean Squared Error (MSE), which measures the difference between the predicted location and the actual location.
[0103] ②Auxiliary loss function:
[0104] - Euclidean distance loss: directly measures spatial distance error
[0105] - Direction loss: Measures the consistency between the predicted direction and the actual direction.
[0106] - Regularization term: L2 regularization, to prevent overfitting.
[0107] (4) Optimization algorithm:
[0108] ①Adam optimizer: Adaptive learning rate, accelerates convergence
[0109] ② Learning rate scheduling: The initial learning rate is 0.001, and it decays to 0.8 times the original rate every 50 epochs.
[0110] ③ Early stopping strategy: If the validation set loss shows no improvement after 15 consecutive epochs, training is stopped. 5. Terminal localization method
[0111] (1) Real-time positioning process:
[0112] ① The drone flies along a preset path and collects the terminal RSSI value in real time.
[0113] ② Preprocess the collected data
[0114] ③ Input the processed data into the trained deep learning model
[0115] ④ Predicted position of the model output terminal
[0116] ⑤ Combining historical prediction results, the final location estimate is obtained through weighted averaging or Kalman filtering.
[0117] (2) Multiple measurement fusion:
[0118] ① Perform multiple measurements on the same terminal from different directions and heights.
[0119] ②Weighted fusion of multiple measurement results
[0120] ③ The weights can be dynamically adjusted based on factors such as RSSI intensity and measurement distance.
[0121] (3) Methods to improve positioning accuracy:
[0122] ① Adaptive flight path: Based on the initial positioning results, adjust the UAV's flight path to obtain more effective data.
[0123] ② Multi-model ensemble: Train multiple deep learning models with different structures and improve localization accuracy through ensemble learning.
[0124] ③ Prior information fusion: Combine prior information such as topographic maps and building distribution to optimize the positioning results.
[0125] 6. System Implementation and Deployment
[0126] (1) Hardware implementation:
[0127] ① Drone platform: Select a multi-rotor drone with long endurance and strong wind resistance.
[0128] ② Base station equipment: Wireless receiver module supporting multiple frequency bands and adjustable transmit power.
[0129] ③ Computing platform: edge computing devices (such as NVIDIA Jetson Nano) or cloud servers
[0130] ④ Communication module: 4G / 5G module or long-range WiFi module
[0131] (2) Software implementation:
[0132] ① Operating System: Linux-based real-time operating system
[0133] ② Deep learning frameworks: TensorFlow or PyTorch
[0134] ③ Model deployment: TensorFlow Lite or ONNXRuntime
[0135] (3) System optimization:
[0136] ① Model quantization: Convert the floating-point model to an int8 model to reduce computational load.
[0137] ② Parallel processing: Data acquisition and processing are performed in parallel.
[0138] ③ Distributed computing: Rationally allocating computing tasks between the UAV and the ground station.
[0139] ④ Energy consumption optimization: Dynamically adjust the allocation of computing resources according to task requirements.
[0140] Application scenarios
[0141] This invention is applicable to the following application scenarios:
[0142] (1) Search and rescue operation: Locate the signal transmitter carried by the distressed person to improve search and rescue efficiency.
[0143] (2) Wildlife research: Locating wild animals equipped with signal transmitters and studying their activity patterns.
[0144] (3) Asset tracking: Locating important assets in logistics and warehousing.
[0145] (4) Wireless sensor networks: Locating sensor nodes distributed over a wide area
[0146] (5) Security monitoring: Locating suspicious signal sources to assist in security protection.
[0147] in conclusion
[0148] This invention proposes a deep learning-based UAV base station RSSI positioning method and system. By using a single UAV carrying a base station and combining it with deep learning technology, high-precision positioning of terminals is achieved. This method overcomes the limitations of traditional multi-base station positioning methods and has advantages such as simple deployment, low cost, high accuracy, and strong adaptability, providing a new technical solution for various terminal positioning applications.
Claims
1. A deep learning-based unmanned aerial vehicle (UAV) base station RSSI localization method, characterized in that, Includes the following steps: Data acquisition steps: Control the drone carrying the base station to fly along the preset flight path, and record the received terminal RSSI values and the corresponding drone GPS coordinates in real time to form an RSSI-location dataset; Data preprocessing steps: The collected RSSI data is filtered and denoised, the GPS coordinates are standardized, and time series features are constructed to extract RSSI variation trends; Model training steps: Construct a hybrid deep learning model containing CNN and LSTM modules, and train it using the preprocessed dataset. The CNN module is used to extract spatial features of RSSI-location data, and the LSTM module is used to capture the temporal features of RSSI changes with time and location. Terminal positioning steps: Input the RSSI data collected by the UAV in real time into the trained deep learning model, output the predicted position coordinates of the terminal, and improve the positioning accuracy by fusing multiple prediction results.
2. The method according to claim 1, characterized in that, The preset flight path includes at least one of a grid-like path, a spiral path, or an adaptive path that is dynamically adjusted based on real-time RSSI values.
3. The method according to claim 1, characterized in that, The data preprocessing steps specifically include: Noise removal of RSSI signals is achieved by using at least one of median filtering, Kalman filtering, or wavelet transform. Extracting multiple feature combinations from time-domain features, frequency-domain features, spatial features, and gradient features; Data augmentation can be achieved by adding Gaussian noise, randomly dropping data, or generating synthetic data based on signal propagation models.
4. The method according to claim 1, characterized in that, The hybrid deep learning model includes: The CNN module contains at least two convolutional layers and pooling layers for extracting spatial features from RSSI-location data; The LSTM module contains at least two LSTM layers and is used to capture the temporal features of RSSI as it varies with time and location. The fully connected module, which includes multiple fully connected layers and Dropout layers, is used to fuse features and map them to terminal location coordinates.
5. The method according to claim 1, characterized in that, The terminal positioning step further includes: Historical prediction results are fused using either weighted averaging or Kalman filtering methods. The drone's flight path is dynamically adjusted based on the initial positioning results to obtain more effective data. Optimize the positioning results by combining prior information such as topographic maps or building distribution.
6. A UAV base station RSSI positioning system implementing the method of any one of claims 1-5, characterized in that, include: Unmanned aerial vehicle (UAV) platforms are used to carry base station equipment and fly along preset paths. Base station equipment, installed on a drone, is used to receive terminal signals and measure RSSI values; The data processing unit is used for preprocessing the collected data and model inference; The model training unit is used for training and optimizing deep learning models. The communication module is used to enable data transmission between the UAV and the ground station.
7. The system according to claim 6, characterized in that, The data processing unit is deployed on the drone or ground station and adopts an edge-cloud collaborative architecture, supporting incremental learning at the edge and transfer learning in the cloud.
8. The system according to claim 6, characterized in that, The base station equipment supports multi-band, adjustable transmit power wireless signal reception, and the data processing unit uses a quantized int8 model for real-time inference.
9. The system according to claim 6, characterized in that, It also includes a multi-model integration module, which integrates the prediction results of multiple deep learning models with different structures to improve positioning accuracy.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-5.