Method for cooperative navigation and communication of low-altitude unmanned aerial vehicle based on visible light positioning

Abstract

The application discloses a kind of based on visible light positioning's middle-low altitude unmanned aerial vehicle cooperative navigation and communication method, method includes middle space reference unmanned aerial vehicle by image matching positioning obtains accurate three-dimensional position of itself, calibrates the absolute space coordinates of each LED in LED array, using frequency division multiplexing control LED array downward emission while carrying the superposition signal of positioning signal and communication signal, form visible light signal;Low altitude unmanned aerial vehicle to be positioned by single photoelectric detector receives visible light signal, sequentially executes signal windowing, FFT transform and RSS vector extraction operation, then denoising by Transform deep learning network, power function attenuation model ranging, combined with PPO reinforcement learning algorithm adaptive output positioning parameter, obtain accurate three-dimensional coordinates of itself by weighted least squares solution;Low altitude unmanned aerial vehicle to be positioned by FFT, Morlet wavelet transform, hilbert transform three-stage joint decoding separation analytical communication signal, combined with forward error correction and CRC check realizes the two-way communication between middle-low altitude unmanned aerial vehicle.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) navigation and positioning technology, and more specifically, relates to a method for cooperative navigation and communication of low- and medium-altitude UAVs based on visible light positioning. Background Technology

[0002] With the rapid development of drone technology, its applications in logistics delivery, environmental monitoring, emergency rescue, urban inspection, and other fields are becoming increasingly widespread. Navigation, positioning, and communication are the core technologies supporting the efficient and safe operation of drones. Currently, drone navigation and positioning mainly rely on the Global Navigation Satellite System (GNSS). However, in satellite-denied environments such as indoor warehouses, underground parking lots, and urban canyons, satellite signals are easily blocked, interfered with, or spoofed, leading to GNSS navigation failure and the inability to provide reliable location information for drones, severely limiting their operational range and application scenarios.

[0003] To address the navigation challenges of drones in satellite-denied environments, several alternative solutions have emerged, but all suffer from significant drawbacks: First, WiFi / Bluetooth fingerprint positioning technology offers low accuracy, typically with an error of 3-5 meters, failing to meet the precision requirements of drone operations. Second, while ultra-wideband (UWB) positioning technology boasts higher accuracy, it necessitates the deployment of multiple fixed anchor points, resulting in high costs, complex wiring, and poor deployment flexibility, making it unsuitable for temporary tasks or dynamic operational scenarios. Third, visual SLAM technology demands substantial computation and is highly sensitive to lighting conditions, prone to positioning drift in low-light, high-light, or abrupt changes in illumination, exhibiting insufficient robustness. Fourth, existing visible light positioning (VLP) systems often employ multi-anchor fixed deployments or complex reception methods. The design of the terminal is complex and difficult to deploy. Furthermore, the fixed positioning algorithm parameters make it difficult to adapt to signal variations in different environments, leading to unstable positioning accuracy. Fifth, ambient light interference, multipath effects, and electronic noise severely affect the measurement accuracy of received signal strength (RSS). Existing technologies have limited signal noise suppression capabilities, further restricting positioning accuracy. Sixth, traditional solutions separate navigation and communication functions, requiring independent communication systems, increasing the hardware complexity, weight, and power consumption of the UAV, which is detrimental to the endurance and integration of small UAVs. Seventh, the lack of an effective cooperative positioning mechanism fails to fully utilize the auxiliary positioning advantages of known location nodes in the UAV swarm, making it difficult to further improve positioning coverage and accuracy.

[0004] Therefore, there is an urgent need to develop a mid-to-low altitude UAV collaborative navigation and communication technology that is flexible in deployment, accurate in positioning, highly adaptable to the environment, and integrates navigation and communication, in order to solve the problem of UAV operation in satellite-denied environments. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for cooperative navigation and communication of low- and mid-altitude unmanned aerial vehicles (UAVs) based on visible light positioning. By employing a Transformer deep learning network, it mines the spatial correlation between the received signal strength (RSS) values ​​of multiple LEDs, effectively suppressing ambient light interference, multipath effects, and electronic noise, thus improving the accuracy of RSS measurement after noise reduction by 15-25%. Simultaneously, it combines a PPO reinforcement learning algorithm to automatically learn optimal positioning parameters, innovatively introducing smoothing penalty terms and over-limit penalty terms to balance instantaneous positioning accuracy, trajectory continuity, and safety. Compared to fixed-parameter algorithms, this improves positioning accuracy by 20-30%. Furthermore, through online learning, it continuously adapts to different environmental conditions, solving the problems of unstable accuracy and poor environmental adaptability in traditional positioning technologies. Drawing inspiration from the bioluminescent communication mechanism of fireflies, it employs a dual encoding method of frequency and intensity. This system employs a three-level joint decoding algorithm combining FFT, wavelet transform, and Hilbert transform, achieving a communication error rate of less than 1%. Simultaneously, it integrates high-frequency positioning signals and low-frequency communication signals via an LED array, eliminating the need for additional communication equipment and reducing system weight and power consumption by 30-50%. This significantly simplifies the hardware structure, lowers costs, and enhances the drone's endurance. Based on a mid-to-low altitude collaborative architecture, the system uses mid-altitude drones as flight positioning base stations, eliminating the need for fixed ground anchors. This allows for rapid deployment and supports multi-drone networking to expand coverage. Low-altitude drones require only a single photodetector to separate signals via frequency division multiplexing. This results in high hardware integration, low cost, and immunity to electromagnetic interference in the visible light band. It can operate stably in satellite-denied environments, electromagnetically dense environments, or electromagnetically silent scenarios, significantly expanding the application boundaries of drones and adapting to various operational needs such as logistics delivery, environmental monitoring, and emergency rescue.

[0006] To achieve the above objectives, one aspect of the present invention provides a method for cooperative navigation and communication of low- and mid-altitude unmanned aerial vehicles (UAVs) based on visible light positioning, comprising the following steps: S1. Positioning and Signal Transmission of the Medium-Altitude Reference UAV: ​​The medium-altitude reference UAV obtains its precise three-dimensional position through image matching positioning. Combined with the three-dimensional position, the absolute spatial coordinates of each LED in the multiple LED arrays carried by the UAV are determined. Frequency division multiplexing technology is used to control the LED arrays carried by the medium-altitude reference UAV to transmit a superimposed signal that simultaneously carries high-frequency positioning signals and low-frequency communication signals downward, forming a visible light signal that can be stably propagated and effectively received within the coverage radius. S2. Low-altitude UAV navigation and positioning calculation: The low-altitude UAV to be positioned receives the visible light signal through a single photodetector on board, and sequentially performs signal windowing, FFT transformation and RSS vector extraction operations. Then, it is optimized by the Transformer deep learning network for noise reduction and distance estimation by the power function decay model. Combined with the PPO reinforcement learning algorithm, the positioning parameters are adaptively output. Finally, the precise three-dimensional coordinates of the UAV are determined by weighted least squares solution. S3. Low-altitude UAV communication decoding: After receiving the visible light signal transmitted in step S1, the low-altitude UAV to be located performs preprocessing, three-level joint decoding and fault-tolerant verification operations in sequence to accurately parse the communication data and realize two-way communication between low- and medium-altitude UAVs for receiving flight control commands, transmitting mission parameters and providing status information feedback.

[0007] Further, in step S1, the use of frequency division multiplexing technology to control the LED array on the mid-altitude reference UAV to emit a superimposed signal carrying both positioning and communication signals downwards, forming a visible light signal with a stable propagation range and effective reception coverage, includes: The relative position of each LED in the array is pre-calibrated based on the precise three-dimensional position of the drone, and the absolute spatial coordinates of each LED are determined. Each LED is assigned multiple non-overlapping modulation frequencies to generate a high-frequency positioning signal in the 200-3000Hz band; at the same time, the low-frequency communication signal in the 1-20Hz band is encoded using a dual modulation method of "different frequencies corresponding to flight commands and brightness changes encoding binary data", and then superimposed on the positioning signal of each LED. The LED array is controlled to emit a superimposed signal that simultaneously carries a high-frequency positioning signal of 200-3000Hz and a low-frequency communication signal of 1-20Hz, forming a visible light signal with a coverage radius of 20-50 meters. The positioning signal and communication signal frequency bands are separated and the frequency interval is more than 100 times.

[0008] Furthermore, step S2 specifically includes: S21. The photodetector acquires the time-domain data sequence of the visible light signal at a preset sampling rate, performs windowing processing on the time-domain data sequence, and performs fast Fourier transform frequency domain analysis to extract the received signal strength RSS vector corresponding to each LED modulation frequency. S22. Input the RSS vector into the Transformer deep learning network, correct outliers by learning the spatial correlation between various LED signals, and output the denoised RSS vector. S23. Based on the power function decay model, the distance estimate from the low-altitude UAV to each LED is calculated using the denoised RSS vector. S24. The positioning parameters are adaptively optimized by using the PPO reinforcement learning optimization algorithm. Combining the distance estimate with the known spatial position of the LED, the three-dimensional coordinates of the low-altitude UAV to be positioned are calculated using the weighted least squares algorithm.

[0009] Further, step S21 includes: S211, the photodetector uses a preset sampling rate Continuously acquire time-domain data sequences of visible light signals; S212, for lengths of The time-domain signal is windowed, and the window size is... The number of windows is ), The total length of the time-domain signal; S213, Execute on the data of each window. Point Fast Fourier Transform (FFT) and calculate the one-sided magnitude spectrum; S214. Determine the frequency resolution based on the sampling rate and window size. The calculation formula is: S215. Calculate the corresponding frequency index for the modulation frequency of each LED. The calculation formula is: S216. Frequency index corresponding to the modulation frequency of each LED The peak value is found within ±2 frequency points nearby as the RSS value, and finally formed 3D RSS vector .

[0010] Furthermore, the structure of the Transformer deep learning network described in step S22 includes: Input layer: RSS vector , dimension is ; Location encoding layer: for position pos and dimension 2 The formula for calculating the location code is: in, The dimension of the Transformer model; This indicates a position of pos and a dimension of 2. The position encoding result at that time; This indicates that the position is pos and the dimension is The position encoding result at that time; Indicates the index of the current dimension; Transformer encoder layer: Stacked 4 Transformer Encoder layers, each layer includes a multi-head self-attention mechanism, a feedforward neural network, layer normalization and residual connections; Output layer: Outputs the denoised RSS vector. The RSS value affected by noise pollution is corrected to a value that conforms to physical laws.

[0011] Furthermore, the expression for the power function decay model described in step S23 is: in, For low-altitude drones to the first The distance between LEDs; For the denoised first The received signal strength of each LED; , All parameters were obtained through calibration experiments. The distance calculation involves substituting the denoised RSS values ​​of each LED into the power function attenuation model to obtain the distance. Several distance estimates are used to form a distance vector. ; Furthermore, step S24 specifically includes: S241. Constructing the state space: at time... Define system state The system state The comprehensive expression, which incorporates information related to current signal characteristics, positioning error, and environmental interference, is as follows: in: The received signal strength of each LED after denoising is output in step S22; The distance estimates from the low-altitude UAV to be located to each LED are calculated in step S23. The standard deviation of the denoised RSS vector; They are time points , The positioning error; SNR is the signal-to-noise ratio of the current visible light signal; S242. Define the action space: the actions output by the PPO policy network. It contains the key parameters of the localization algorithm, expressed as: in: For the first The weighting coefficients of each LED satisfy the constraints. ( =1,…, ); For regularization parameters; To optimize the step size factor; S243. To achieve synergistic optimization of positioning accuracy, trajectory smoothness, and positioning security, a reward function is designed. The expression is: in: This represents the positioning error at the current moment. This is a smoothing penalty term for the positioning trajectory; This is a penalty item for exceeding the limit; , All are preset positive numbers, which are the weight coefficients of the smoothing penalty term and the over-limit penalty term, respectively; S244, PPO Strategy Update: The Actor network, employing an Actor-Critic architecture, outputs localization parameters, while the Critic network evaluates the current system state. The value of this is introduced by the pruning objective function. The policy update range is limited; the policy parameters of the Actor network are updated by maximizing the pruning objective function, and the generalized advantage estimation (GAE) method is combined to stably estimate the advantage function. S245. Based on the positioning parameters output by the PPO policy network after the update in step S244, and combined with the distance estimates obtained in step S23 and the known absolute spatial coordinates of each LED, a weighted residual function is constructed; using the three-dimensional position of the UAV to be positioned as the optimization variable, ... The optimization objective is to minimize the sum of the weighted residual squares corresponding to each LED and the sum of the position regularization term. The nonlinear least squares method is used for iterative solution to obtain the accurate three-dimensional coordinates of the UAV to be located. S246. The three-dimensional coordinate positioning result of the low-altitude UAV obtained in step S245 is directly output to the UAV flight control system to realize the real-time navigation control of the UAV. At the same time, the positioning error between the current positioning result and the actual position is calculated, and the positioning error is fed back to the PPO reinforcement learning network as a feedback signal for the network's online learning, so as to realize the dynamic iteration of positioning parameters and the continuous improvement of positioning accuracy.

[0012] Furthermore, in step S244, the objective function is pruned. The expression is: in: Importance sampling ratio: This is the clipping function; The dominant function; The weighted residual function expression in step S245 is: in, For the first The weighted residuals corresponding to each LED; For the drone to be located to the first The geometric distance between the LEDs; Distance estimated based on RSS; For the first The weighting coefficient of each LED; The objective function expression is: in, To find the optimal drone location This minimizes the value of the subsequent function. express The sum of squares of the weighted residuals of each LED; Represents a regular term; These are the regularization parameters output by PPO. It's the location of the drone. It's an initial position guess.

[0013] Furthermore, the three-level decoding in step S3 includes: First-level decoding extracts the low-frequency spectral components of the visible light signal through fast Fourier transform; The second-level decoding uses Morlet wavelets as the basis function to perform continuous wavelet transform, calculates wavelet coefficients, and determines the occurrence time and switching time of significant frequency components through the time-frequency distribution of wavelet coefficients, captures dynamic command sequences, and identifies the amplitude modulation time pattern. Three-level decoding: The instantaneous amplitude, instantaneous frequency, and instantaneous phase of the communication signal are extracted through Hilbert transform, so as to achieve accurate analysis of the dual-modulated signal.

[0014] The second aspect of the present invention provides a mid-to-low altitude UAV cooperative navigation and communication system based on visible light positioning, for implementing the aforementioned mid-to-low altitude UAV cooperative navigation and communication method based on visible light positioning, comprising: a mid-altitude reference UAV signal transmission module, a low-altitude UAV navigation and positioning module, a low-altitude UAV communication decoding module, and a mid-to-low altitude cooperative control module. A mid-altitude reference UAV signal transmission module: mounted on a mid-altitude reference UAV, comprising a positioning submodule, an LED array modulation submodule, and a communication coding submodule; the positioning submodule is used to obtain the precise three-dimensional position of the mid-altitude reference UAV through image matching positioning; the LED array modulation submodule is used to carry multiple LEDs in a preset layout, calibrate the absolute spatial coordinates of each LED according to the three-dimensional position obtained by the positioning submodule, generate a high-frequency positioning signal using frequency division multiplexing technology, and drive the LED array to transmit; the communication coding submodule is used to generate a low-frequency communication signal and perform dual encoding, superimposing it on the positioning signal; Low-altitude UAV navigation and positioning module: Equipped on a low-altitude UAV to be positioned, it includes a photodetector submodule, a signal processing submodule, a Transformer denoising submodule, a distance estimation submodule, and a PPO reinforcement learning positioning submodule. The photodetector submodule receives visible light signals emitted by the signal transmitting module of a mid-altitude reference UAV. The signal processing submodule performs windowing, FFT transformation, and RSS vector extraction on the received signal. The Transformer denoising submodule optimizes the RSS vector for denoising. The distance estimation submodule calculates the distance estimate based on a power function decay model. The PPO reinforcement learning positioning submodule adaptively outputs positioning parameters and calculates the three-dimensional coordinates of the low-altitude UAV using weighted least squares. Low-altitude UAV communication decoding module: Equipped on a low-altitude UAV awaiting positioning, it includes a frequency band separation submodule, a three-level joint decoding submodule, and a fault tolerance verification submodule. The frequency band separation submodule is used to separate the high-frequency positioning signal and the low-frequency communication signal from the visible light signal. The three-level joint decoding submodule is used to sequentially analyze the frequency domain, time-frequency, and instantaneous parameters of the low-frequency communication signal using Fourier transform, wavelet transform, and Hilbert transform. The fault tolerance verification submodule is used to ensure communication reliability and restore communication data through forward error correction codes and cyclic redundancy check. Low- and medium-altitude collaborative control module: used to realize vertical collaboration between low-altitude reference UAVs and low-altitude UAVs to be positioned, network collaboration of multiple low-altitude reference UAVs, and parallel scheduling of navigation and communication functions, including base station scheduling of low-altitude reference UAVs, temporary anchor point management of positioned low-altitude UAVs, and parallel scheduling of positioning and communication functions.

[0015] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) This invention uses a Transformer deep learning network to mine the spatial correlation between the received signal strength (RSS) values ​​of multiple LEDs, which can effectively suppress ambient light interference, multipath effect and electronic noise. Compared with traditional filtering methods, it can adapt to complex noise patterns and improve the RSS measurement accuracy after denoising by 15-25%, providing accurate data support for distance estimation. At the same time, the optimal positioning parameters such as LED weights and regularization parameters are automatically learned through the PPO reinforcement learning algorithm. Furthermore, a smoothing penalty term and an over-limit penalty term are innovatively introduced in the reward function, which not only ensures the instantaneous positioning accuracy, but also takes into account the trajectory continuity and safety. Experimental verification shows that the positioning accuracy is improved by 20-30% compared with the fixed parameter algorithm. Moreover, with the help of online learning capabilities, it can continuously adapt to different environmental conditions, solving the technical problems of fixed parameters and poor environmental adaptability of traditional positioning algorithms.

[0016] (2) This invention draws on the bioluminescent communication mechanism of fireflies and uses a dual encoding method combining frequency modulation and intensity modulation to transmit information. It naturally has strong anti-interference capabilities, and can increase information capacity and improve transmission efficiency. At the same time, it also has the concealment of biological communication, which is suitable for electromagnetic silence requirements. The LED array simultaneously undertakes the functions of high-frequency band (e.g., 200-3000Hz) positioning signal transmission and low-frequency band (e.g., 1-20Hz) communication data transmission, realizing hardware sharing. No additional communication equipment is required. Compared with the traditional separate solution, the system weight and power consumption are reduced by 30-50%, effectively improving the endurance of UAVs and making it suitable for small UAV applications. In addition, the three-level joint decoding algorithm of FFT, wavelet transform and Hilbert transform is adopted to fully integrate frequency domain, time-frequency domain and instantaneous parameter information. The three complement each other and enhance each other. Combined with forward error correction code and cyclic redundancy check mechanism, the communication error rate is less than 1%. It can still work stably even in strong noise environment and ensure communication reliability.

[0017] (3) The present invention adopts a mid-to-low altitude vertical collaborative architecture, using mid-altitude UAVs as flight positioning base stations, breaking through the limitations of traditional fixed anchor points, and can be quickly deployed to adapt to temporary tasks and dynamic scenarios. Multiple mid-altitude UAVs can be networked to expand the coverage area. Its geometric configuration has better positioning accuracy and higher line-of-sight communication probability than ground multi-anchor point systems, and can reduce the impact of occlusion. Low-altitude UAVs only need a single photoelectric detector, and multiple LED signals are separated in the frequency domain through frequency division multiplexing technology. Compared with existing visible light positioning (VLP) systems that require PD arrays or cameras, the hardware structure is greatly simplified, the cost is reduced, and it is easy to integrate with small UAVs. At the same time, the visible light frequency band is independent of the radio frequency band and is not affected by electromagnetic interference from WiFi, Bluetooth, mobile phone signals, radar, etc. It can work normally in electromagnetically dense environments or electromagnetically silent conditions, significantly expanding the application boundaries of UAVs in satellite-denied environments and adapting to various scenarios such as urban canyons, indoor warehouses, and emergency rescue. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning, according to an embodiment of the present invention. Figure 2 This is a logical architecture diagram of low- and mid-altitude collaborative positioning in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Transformer deep learning network structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the PPO reinforcement learning-optimized localization process in an embodiment of the present invention; Figure 5 This is a flowchart of the three-level decoding of communication signals in an embodiment of the present invention; Figure 6This is a schematic diagram of a low-to-medium altitude unmanned aerial vehicle (UAV) cooperative navigation and communication system based on visible light positioning, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] In environments where GNSS signals are unavailable, such as urban canyons, indoor warehouses, and underground parking lots, traditional navigation systems fail, and drones cannot obtain precise locations. Based on the above reasons, such as... Figure 1 and Figure 2 As shown, one aspect of the present invention provides a cooperative navigation and communication method for low- and mid-altitude unmanned aerial vehicles (UAVs) based on visible light positioning, mainly applied to low-altitude UAV formation operation scenarios in satellite-denied environments, including the following steps: S1. Mid-altitude reference UAV positioning and signal transmission: The mid-altitude reference UAV obtains its precise three-dimensional position through image matching positioning. Combined with the three-dimensional position, the absolute spatial coordinates of each LED in the multiple LED arrays carried by the UAV are determined. Frequency division multiplexing technology is used to control the LED arrays carried by the mid-altitude reference UAV to transmit a superimposed signal that simultaneously carries high-frequency (e.g., 200-3000Hz) positioning signal and low-frequency communication signal, forming a visible light signal that can be stably propagated and effectively received within the coverage radius. S2. Low-altitude UAV navigation and positioning calculation: The low-altitude UAV to be positioned receives the visible light signal through a single photodetector on board, and sequentially performs signal windowing, FFT transformation and RSS vector extraction operations. Then, it is optimized by the Transformer deep learning network for noise reduction and distance estimation by the power function decay model. Combined with the PPO reinforcement learning algorithm, the positioning parameters are adaptively output. Finally, the precise three-dimensional coordinates of the UAV are determined by weighted least squares solution. S3. Low-altitude UAV communication decoding: After receiving the visible light signal transmitted in step S1, the low-altitude UAV to be located performs preprocessing, three-level joint decoding and fault-tolerant verification operations in sequence to accurately parse the communication data and realize two-way communication between low- and medium-altitude UAVs for receiving flight control commands, transmitting mission parameters and providing status information feedback.

[0021] Furthermore, the hollow reference UAV described in step S1 obtains its precise three-dimensional position through image matching and positioning. The image matching and positioning step includes: Real-time ground images of the work area are captured using a visible light camera mounted on a mid-altitude reference drone. Retrieve pre-stored satellite remote sensing images with clear geographical location information as a reference benchmark; use feature point matching algorithms (such as SIFT, ORB algorithms) or deep learning image matching models to perform matching operations between real-time captured ground images and pre-stored satellite remote sensing images; The precise three-dimensional coordinates of the mid-altitude reference UAV were determined by solving the matching results. ; Furthermore, after obtaining the position in step S1, the mid-altitude UAV is equipped with an LED array transmitter; the LED array includes multiple LEDs arranged in a preset geometric layout, the relative position of each LED is known, and each LED uses a different modulation frequency (e.g., in one embodiment of the present invention, there are 8 LEDs, and the modulation frequencies are 208Hz, 354Hz, 502Hz, 662Hz, 678Hz, 826Hz, 906Hz, and 922Hz respectively). The method employs frequency division multiplexing (FDM) technology to control the LED array onboard the mid-altitude reference UAV to emit a superimposed signal simultaneously carrying positioning and communication signals downwards, forming a visible light signal, including: The relative position of each LED in the array is pre-calibrated based on the precise three-dimensional position of the drone, and the absolute spatial coordinates of each LED are determined. Each LED is assigned multiple non-overlapping modulation frequencies. ( =1,2,… The system generates a high-frequency positioning signal (in the embodiment of the invention, the high-frequency band refers to 200-3000Hz) to achieve frequency division multiplexing of the positioning signal; at the same time, the low-frequency communication signal (in the embodiment of the invention, the low-frequency band refers to 1-20Hz) is encoded using a dual modulation method of "different frequencies correspond to flight commands, and brightness changes are encoded into binary data" and then superimposed on the positioning signal (driving current) of each LED. The LED array is controlled to emit a superimposed signal that simultaneously carries high-frequency positioning signals (e.g., 200-3000Hz) and low-frequency communication signals (e.g., 1-20Hz), forming a visible light signal with a coverage radius of 20-50 meters. The positioning signal and the communication signal are separated by frequency bands and the frequency interval is more than 100 times, so they do not interfere with each other.

[0022] Furthermore, step S2 specifically includes: S21. Signal reception and RSS extraction: The photodetector acquires the time-domain data sequence of the visible light signal at a preset sampling rate. After windowing the time-domain data sequence, it performs Fast Fourier Transform (FFT) frequency domain analysis to extract the received signal strength RSS vector corresponding to each LED modulation frequency. In a specific embodiment of the present invention: the photodetector uses a sampling rate Optical signals were continuously acquired at 2000Hz to obtain a time-domain data sequence. The acquired data was then windowed with a window size of N=1024. A 1024-point FFT was performed on each window to calculate the single-sided amplitude spectrum. The peak value was found within ±2 frequency points around the 8 LED modulation frequencies to obtain the RSS vector. Suppose that the extracted RSS value at a certain moment is: [0.82, 0.65, 0.48, 0.35, 0.33, 0.25, 0.18, 0.15]; S22, Transformer deep learning denoising: The RSS vector is input into the Transformer deep learning network, and outliers are corrected by learning the spatial correlation between various LED signals, and the denoised RSS vector is output. S23. Distance estimation: Based on the power function decay model, the distance estimate from the low-altitude UAV to be located to each LED is calculated using the denoised RSS vector; S24. PPO reinforcement learning optimized positioning: The positioning parameters are adaptively optimized through the PPO reinforcement learning optimized positioning algorithm. Combined with the distance estimate and the known spatial position of the LED, the three-dimensional coordinates of the low-altitude UAV to be positioned are calculated using the weighted least squares algorithm.

[0023] Further, step S21 includes: S211, the photodetector uses a preset sampling rate Continuously acquire time-domain data sequences of visible light signals; S212, for lengths of The time-domain signal is windowed, and the window size is... The number of windows is ), The total length of the time-domain signal; S213, Execute on the data of each window. Point-wise Fast Fourier Transform (FFT) is performed, and the one-sided magnitude spectrum is calculated; the formula for calculating the Fast Fourier Transform is as follows: in, The result of the FFT transformation; The time-domain signal data within a single window to be processed; This represents the number of transformation points in the FFT, corresponding to the window size. The formula for calculating the single-sided amplitude spectrum is: in, The final one-sided amplitude spectrum (only retaining values ​​from 0 to 1) Amplitude within the frequency range, and through (Normalize the amplitude so that the frequency domain amplitude corresponds to the energy of the original time domain signal). S214. Determine the frequency resolution based on the sampling rate and window size. The calculation formula is: S215. Calculate the corresponding frequency index for the modulation frequency of each LED. The calculation formula is: S216. Frequency index corresponding to the modulation frequency of each LED Find the peak value within ±2 frequency points nearby as the RSS value: In a specific embodiment of the present invention, the final form is 3D RSS vector: .

[0024] Furthermore, in step S22, the RSS vector extracted in the previous step is input into the Transformer attention mechanism network for denoising to suppress ambient light interference, multipath effects and electronic noise; the Transformer deep learning network is a deep learning model based on a self-attention mechanism. The structure of the Transformer deep learning network includes: Input layer: RSS vector , dimension is ; Position Encoding Layer: Since the Transformer itself does not have the ability to perceive position information, position encoding needs to be added; for position pos and dimension 2 The formula for calculating the location code is: in, The dimension of the Transformer model; This indicates a position of pos and a dimension of 2. The positional encoding result at that time corresponds to a sine code, used for even-numbered dimensions; This indicates that the position is pos and the dimension is The positional encoding result at time corresponds to cosine encoding, used for odd-numbered dimensions; Indicates the index of the current dimension; Transformer encoder layer: Stacked 4 Transformer Encoder layers, each layer includes a multi-head self-attention mechanism, a feed-forward network, layer normalization and residual connections; Output layer: Outputs the denoised RSS vector. Correct outliers caused by noise pollution; The scaling dot product attention expression is: ,in, (Query) represents the "query" matrix, which represents the information to be matched in the current input; (Key) represents the "key" matrix, which identifies the candidate information; (Value) represents the "value" matrix, which represents the actual content of the candidate information; express The dimension of the matrix (corresponding to "key vector dimension set to 16"). Indicates calculation and Similarity (dot product); Indicates passage Scaling to avoid the dot product result being too large. Gradient vanishing; This represents converting similarity into attention weights (with a weight sum of 1); the final output is the attention weights and their sums. The weighted sum, which incorporates the information of interest; The multi-head attention expression is: in, express One independent "scaled dot product attention" result (corresponding to "number of attention heads"); Indicates multiple The results are concatenated into a matrix; This represents the output projection matrix, which maps the concatenated matrix to the target dimension; the multi-head mechanism enables the model to focus on information from different representation subspaces at different locations; For the first The output of each attention head; This is a scaled dot product attention function; This represents the original query matrix (the original input to the multi-head attention module). ); This represents the original key matrix (the original key matrix input to the multi-head attention module). ); This represents the original value matrix (the original input to the multi-head attention module). ); Indicates the first Each attention head corresponds to Projection matrix; Indicates the first Each attention head corresponds to Projection matrix; Indicates the first Each attention head corresponds to Projection matrix; The layer normalization expression is: , in, The input data to be normalized; for The mean, ; for variance ; To prevent small constants from being divided by zero, i.e., local minima (to avoid denominators of 0). ; These are the learnable scaling and offset parameters, respectively. This represents element-wise multiplication; by stabilizing the network training through layer normalization, convergence is accelerated and internal covariate bias is reduced; The expression for a feed-forward network is: ,in, This is the output of the feedforward neural network; This is the input data for the feedforward neural network; This is the weight matrix of the first fully connected layer; This is the bias vector of the first fully connected layer; This is the weight matrix for the second fully connected layer; This is the bias vector of the first fully connected layer; Use the ReLU activation function; The residual connection expression is: in, This serves as the input for sub-layers (such as attention layers / feedforward layers); For the output of sub-layers (such as attention layers / feedforward layers); This indicates that the result of adding the residuals is normalized by layer; residual connections make the network easier to train while preserving the information of the original input.

[0025] The denoising principle of step S22 is as follows: Transformer learns through a self-attention mechanism. The spatial correlation between the RSS values ​​of individual LEDs; under normal circumstances, the RSS values ​​of LEDs that are close to each other should show a specific relative relationship, but noise disrupts this relationship; the Transformer network learns the normal pattern of RSS vectors through training, and corrects the noise-contaminated RSS values ​​to values ​​that conform to physical laws, thereby achieving noise reduction. In embodiments of the present invention, such as Figure 3 As shown, the RSS vector is input into a Transformer network for denoising. The network structure is: input layer (8 dimensions) → position encoding → Transformer Encoder × 4 layers → output layer (8 dimensions). Each layer includes a multi-head self-attention layer (4 attention heads), residual connections, layer normalization, and a feedforward network. The denoised RSS value becomes: [0.80, 0.66, 0.50, 0.36, 0.34, 0.24, 0.17, 0.14], and the noise is effectively suppressed. Furthermore, step S23, based on the physical law of light intensity attenuation with distance, uses a pre-calibrated power function attenuation model to convert the denoised RSS value into an estimated distance from the low-altitude UAV to each LED; the expression for the power function attenuation model is: in, For low-altitude drones to the first The distance between LEDs (unit: meters); For the denoised first The received signal strength of each LED; , All parameters were obtained through calibration experiments (each LED was fitted independently). The typical value is approximately -3, which is negative and indicates that the distance increases as the RSS decreases. Parameters The fitting process includes: collecting RSS data at known locations, calculating the true geometric distance, and fitting the power function parameters using the nonlinear least squares method (Levenberg-Marquardt algorithm) to minimize the sum of squared residuals between the fitted distance and the true distance; Distance calculation: Substitute the denoised RSS values ​​of each LED into the model to calculate the distance. Several distance estimates are used to form a distance vector. ; In an embodiment of the present invention, it is assumed that the parameters of the first LED are obtained through a calibration experiment: ,but: Repeat the calculation for 8 LEDs to obtain the distance vector: [3.9, 4.8, 6.2, 7.8, 8.1, 9.5, 11.2, 12.5] meters; Furthermore, step S24 specifically includes: S241. Constructing the state space: at time... Define system state The system state The comprehensive expression, which incorporates information related to current signal characteristics, positioning error, and environmental interference, is as follows: in: The received signal strength of each LED after denoising is output in step S22; The distance estimates from the low-altitude UAV to be located to each LED are calculated in step S23. The standard deviation of the denoised RSS vector is used to characterize the stability of the signal strength. They are time points , The positioning error is used to reflect the historical trend of the positioning trajectory; SNR is the signal-to-noise ratio of the current visible light signal, used to characterize the degree of environmental interference in signal transmission. S242. Define the action space: the actions output by the PPO policy network. It contains the key parameters of the localization algorithm, expressed as: in: For the first The weighting coefficients of each LED satisfy the constraints. ( =1,…, ); This is a regularization parameter used to balance positioning accuracy and solution stability. ; To optimize the step size factor, ; S243. Designing the Reward Function: The reward function is the core of reinforcement learning, determining the direction of policy optimization. This invention designs a reward function to achieve synergistic optimization of positioning accuracy, trajectory smoothness, and positioning security. The expression is: in: This represents the positioning error at the current moment, used to measure the accuracy of the positioning result. The smaller the value, the higher the positioning accuracy. This is a smoothing penalty term for the positioning trajectory, used to constrain abrupt changes in positioning coordinates between adjacent time points and ensure trajectory continuity; This is an over-limit penalty item used to limit the positioning results from deviating from a preset reasonable range and improve positioning security; , All are preset positive numbers, which are the weight coefficients of the smoothing penalty term and the over-limit penalty term, respectively, used to adjust the influence weight of the two penalties in the reward function; Positioning error The primary optimization objective is to minimize the error, resulting in a higher reward. The expression is as follows: By applying a smoothing penalty to mitigate drastic changes in error between adjacent time steps, the positioning results are prevented from fluctuating, resulting in a smooth trajectory. The expression is as follows: When the positioning error exceeds a safety threshold (e.g., 2 meters), an additional penalty is applied to guide the strategy and prevent large errors; the over-limit penalty expression is: In an embodiment of the present invention, the weighting coefficients are determined through experimental optimization: ; S244, PPO Policy Update: A dual-network architecture of Actor-Critic is used to implement policy updates, where the Actor network is used to update the system status based on step S241. Output the positioning parameters (motion vectors) in step S242 The Critic network is used to evaluate the current system state. The value of the policy update; to avoid excessive policy update magnitude leading to training instability, a pruning objective function is introduced. The policy update range is limited, and the policy parameters of the Actor network are updated by maximizing the pruning objective function. At the same time, the generalized advantage estimation (GAE) method is combined to stably estimate the advantage function and update the parameters of the Actor network and the Critic network synchronously. In the Actor-Critic architecture, the Actor network and the Critic network are trained in parallel. The Actor network adjusts the positioning parameters and outputs the policy according to the advantage function estimated by the Critic network. The Critic network corrects the state value estimation through temporal difference (TD) error, thereby achieving adaptive optimization of the positioning parameters. In embodiments of the present invention, the Actor network (policy network) includes: Input layer: State vector Hidden layers: 3 fully connected layers with [256, 128, 64] neurons each, and ReLU activation function. Output layer: LED weights After softmax normalization, the sum is guaranteed to be 1; parameters After being mapped to the corresponding range using a sigmoid function, the final output is the action vector. (Dimension 10); Critic Network (Value Network): Input: State vector Hidden layers: 3 fully connected layers with [256, 128, 64] neurons each, and ReLU activation function. Output: State value (Scalar); Pruning objective function The expression is: in: Importance sampling ratio: This is the clipping function; The dominant function; The expression for the importance sampling ratio is: This indicates a new strategy. Compared to the old strategy The ratio; The pruning function is used to limit the importance sampling ratio to a certain value. Within the range, the expression is: ,in, As the cropping threshold, in this invention ; Advantage function Calculated using generalized dominance estimation (GAE): in, For the first The advantage function of time measures the action. The "goodness" or "badness" relative to the current strategy; For time-difference (TD) residuals; For the first Instant rewards obtained after performing an action; For state State value; For the next state State value; As a discount factor, ; For GAE attenuation parameters, ; The state value function is estimated by the Critic network; S245, 3D coordinate calculation: Based on the positioning parameters output by the PPO strategy network after the update in step S244 ( , Combined with the distance estimate obtained in step S23 Given the known absolute spatial coordinates of each LED, a weighted residual function is constructed; based on the three-dimensional position of the UAV to be located... To optimize variables, with The optimization objective is to minimize the sum of the weighted residual squares corresponding to each LED and the sum of the position regularization term. The nonlinear least squares method (Levenberg-Marquardt algorithm) is used for iterative solution to obtain the accurate three-dimensional coordinates of the UAV to be located, thus completing the navigation and positioning calculation. The weighted residual function expression is: ,in, For the first The weighted residuals corresponding to each LED (measuring the difference between the geometric distance and the RSS estimated distance). For the drone to be located to the first Geometric distance between LEDs (theoretical real spatial distance); Distance estimated based on RSS; For the first The weight coefficients of each LED (output by PPO reinforcement learning, used to adjust the localization contribution of different LEDs). The drone to be located Geometric distance of each LED The expression is: in, The three-dimensional position of the low-altitude UAV to be solved (the target parameters to be determined in the end); For the first The known absolute spatial coordinates of each LED (pre-calibrated by the position of the mid-altitude reference UAV and the LED array layout); The objective function expression is: ,in, To find the optimal drone location This minimizes the value of the subsequent function. express The sum of squares of the weighted residuals of each LED; Represents a regular term; It is the regularization parameter output by PPO; It's the location of the drone. It is an initial position guess; S246. Positioning Feedback and Continuous Network Learning: The three-dimensional coordinate positioning result of the low-altitude UAV obtained in step S245 is directly output to the UAV flight control system to realize real-time navigation control of the UAV; at the same time, the positioning error between the current positioning result and the actual position is calculated, and the positioning error is fed back to the PPO reinforcement learning network as a feedback signal for the network's online learning, so as to realize the dynamic iteration of positioning parameters and the continuous improvement of positioning accuracy. The PPO reinforcement learning network adopts a hybrid training mode of "offline pre-training + online continuous learning". In the offline pre-training stage, a large amount of training data under different environmental interference and occlusion conditions is generated in a simulation environment to train the PPO network to learn a general positioning optimization strategy and ensure the initial performance of the network. In the online learning stage, during the actual flight of the UAV, the PPO policy network parameters are continuously updated based on the positioning error feedback in the real scene, so that the network can adaptively match the current flight environment and further improve the environmental adaptability and stability of positioning accuracy. Traditional positioning algorithms use fixed weights or simple weight allocation based on signal quality, while this invention automatically learns the optimal weight allocation strategy through PPO reinforcement learning. By introducing smoothing penalty terms and over-limit penalty terms into the reward function, the strategy not only pursues instantaneous accuracy but also takes into account trajectory continuity and safety. As flight time increases, the PPO strategy becomes more and more accurate, the positioning accuracy continues to improve, and the adaptive learning capability is strong.

[0026] In embodiments of the present invention, such as Figure 4 As shown, the PPO policy network outputs optimal positioning parameters based on the current state; constructs a state vector containing 8 RSS values, 8 distance estimates, RSS standard deviation, historical error, and signal-to-noise ratio, totaling 19 dimensions; PPOActor network inference: the state vector is input into the Actor network, outputting LED weights, regularization parameters, and step size factor; weighted positioning solution: using the weight parameters output by PPO, the Levenberg-Marquardt algorithm is used to solve for the UAV's 3D position; calculate reward and update policy: calculate reward based on positioning error and update the PPO policy; assuming the solved position is (4.8, 1.5, 0.85) meters, the true position is (5.0, 1.6, 0.85) meters, and the positioning error is 0.22 meters; Furthermore, step S3 specifically includes: S31. Signal Preprocessing and Frequency Band Separation: After the photoelectric detector of the low-altitude UAV being positioned receives the visible light signal superimposed with positioning and communication signals, it first performs preprocessing and frequency band separation operations to provide a clean and continuous signal foundation for the subsequent three-level decoding, including: S311, Signal Synchronization: Based on the frequency characteristics of the high-frequency positioning signal (200-3000Hz) in step S1, time synchronization calibration is achieved to determine the time reference for signal sampling and decoding, ensuring the timing consistency of the subsequent decoding process; S312, Frequency Band Separation: Utilizing the characteristic that the frequency bands of high-frequency positioning signals (200-3000Hz) and low-frequency communication signals (1-20Hz) are more than 100 times apart and do not overlap, the received visible light signals are separated by a filter to extract the low-frequency communication signals separately, effectively avoiding interference from high-frequency positioning signals to the communication decoding process; S313, Signal buffer: The low-frequency communication signal obtained after frequency band separation is buffered to form a continuous and stable signal data sequence, providing continuous signal input for subsequent three-level joint decoding operations; S32, Level 1 Decoding: Extracting the spectral components of the low-frequency band (1-20Hz) of the visible light signal through Fast Fourier Transform (FFT); FFT can accurately extract communication frequency components, but it lacks time positioning capability and cannot capture instantaneous frequency changes; Step S32 specifically includes: The separated low-frequency communication signal is windowed, and the window size is set. ; Execute signals for each window Point Fast Fourier Transform yields the frequency domain signal. Frequency domain signal The expression is: ; Based on sampling rate Calculate the frequency domain signal based on the window size. Corresponding frequency resolution The expression is: ; Calculate the one-sided amplitude spectrum of the frequency domain signal: Identify significant frequency peaks within the 1-20Hz frequency band to initially pinpoint the frequency components corresponding to communication commands; S33, Second-level decoding: Continuous wavelet transform is performed using Morlet wavelets as basis functions to calculate wavelet coefficients. The time-frequency distribution of these coefficients determines the occurrence time and switching moments of significant frequency components, capturing dynamic command sequences and identifying the amplitude modulation time pattern. Specifically: The Morlet wavelet expression is: This wavelet combines the time localization characteristics of a Gaussian window with the frequency characteristics of a sinusoidal oscillation, making it suitable for capturing the time-frequency changes of dynamic signals. Use a Gaussian window; It is a sinusoidal oscillation term; The formula for continuous wavelet transform is: in, The wavelet coefficients are those of the continuous wavelet transform. For scale parameters, Used to control the scaling of wavelet bases: The smaller the wavelet, the more "contracted" it is, corresponding to the high-frequency components of the analyzed signal; The larger the wavelet, the more "stretched" it becomes, corresponding to the analysis of low-frequency components, and its behavior is inversely proportional to the frequency. ; These are translation parameters used to control the time position of the wavelet; The original signal to be analyzed; These are wavelet basis functions; for The complex conjugate of the signal is used to calculate the inner product of the signal and the wavelet basis, and to measure the similarity between the two. This is a normalization factor used to ensure consistent energy of wavelet transform at different scales; The system captures instantaneous frequency changes in communication signals (e.g., switching from 5Hz to 12Hz) through two-level decoding, identifies the time pattern of amplitude modulation (e.g., "bright-dark-bright" corresponds to "1-0-1"), and extracts weak communication signals in noisy environments. S34. Three-level decoding: The instantaneous amplitude, instantaneous frequency, and instantaneous phase of the communication signal are extracted using Hilbert transform to complete three-level decoding, achieving accurate analysis of the dual-modulated signal; specifically including: Perform a Hilbert transform on the signal after wavelet transform processing. The transform formula is: in, Original signal The output signal after Hilbert transform; PV represents the Cauchy principal value integral (used to process the integral). The problem of singularities in time); This is the signal to be processed (here, it is the communication signal after wavelet transform). It is an integral variable (representing the integral element of the time dimension); This refers to the current time (the output time of the Hilbert transform). The coefficients of the transformation; The analytic signal is constructed using the Hilbert transform, and its expression is: in, To analyze the signal; The imaginary unit; Extracting instantaneous amplitude from analytic signals Instantaneous phase Instantaneous frequency ; The instantaneous amplitude (envelope) is used for intensity modulation decoding to identify "bright" (logic 1) and "dark" (logic 0), and its expression is: The instantaneous phase expression is: ,in, To determine the argument of the analytical signal; This indicates that the phase is calculated using the ratio of the imaginary part to the real part; Instantaneous frequency is used for frequency modulation decoding to accurately identify frequency modulation-coded flight commands. The expression is: , The time derivative representing the instantaneous phase; S35: Combining the frequency, time-frequency, and instantaneous parameter results obtained from the three-level decoding in steps S32-S34, a decoding decision is made. Simultaneously, a fault-tolerance mechanism is used to reduce the bit error rate and ensure communication reliability. Specifically, this includes: Frequency modulation command decoding: If the instantaneous frequency Falling on the target command frequency The tolerance range, and instantaneous amplitude Greater than the set threshold If so, it is identified as the corresponding flight command; the expression is: in, ; Intensity Modulation Binary Decoding: Set a preset symbol period to determine the instantaneous amplitude. Sampling is performed according to a preset symbol period. After sampling, high amplitude values ​​are translated as logic 1, and low amplitude values ​​are translated as logic 0. Instantaneous amplitude is detected. The periodic changes restore the binary data stream of task parameters, environmental information, etc. Fault tolerance and verification: Forward error correction codes (such as Reed-Solomon codes) are used to correct transmission errors that occur during the decoding process; at the same time, the integrity of the decoded data is verified through the Cyclic Redundancy Check (CRC) mechanism to ensure that the communication error rate is less than 1%, which meets the communication reliability requirements of UAV collaborative operations. Data output: The decoded flight control commands, mission parameters, and other data are transmitted to the UAV flight control system, while the low-altitude UAV's own status information is fed back to the medium-altitude reference UAV, realizing two-way reliable communication between the low-altitude and medium-altitude UAVs; In embodiments of the present invention, such as Figure 5 As shown, the low-altitude UAV performs three-level decoding on the received signal: Level 1 Decoding (FFT): Perform a 2048-point FFT on the signal to extract the low-frequency spectrum from 1-20Hz. Assume a significant peak is detected at 5Hz. Second-level decoding (wavelet transform): Continuous wavelet transform is performed using Morlet wavelets to confirm that the 5Hz signal appears at 0.5 seconds. Three-level decoding (Hilbert transform): Extracting instantaneous frequencies and instantaneous amplitude Within a time window of 0.5-1.5 seconds, Hz, identified as a "forward" command. Simultaneously detect... The periodic changes are decoded into binary data.

[0027] like Figure 6 As shown, a second aspect of the present invention provides a mid-to-low altitude unmanned aerial vehicle (UAV) cooperative navigation and communication system based on visible light positioning, comprising: a mid-altitude reference UAV signal transmission module, a low-altitude UAV navigation and positioning module, a low-altitude UAV communication decoding module, and a mid-to-low altitude cooperative control module. A mid-altitude reference UAV signal transmission module: mounted on a mid-altitude reference UAV, comprising a positioning submodule, an LED array modulation submodule, and a communication coding submodule; the positioning submodule is used to obtain the precise three-dimensional position of the mid-altitude reference UAV through image matching positioning; the LED array modulation submodule is used to carry multiple LEDs in a preset layout, calibrate the absolute spatial coordinates of each LED according to the three-dimensional position obtained by the positioning submodule, generate a high-frequency positioning signal using frequency division multiplexing technology, and drive the LED array to transmit; the communication coding submodule is used to generate a low-frequency communication signal and perform dual encoding, superimposing it on the positioning signal; Low-altitude UAV navigation and positioning module: Equipped on a low-altitude UAV to be positioned, it includes a photodetector submodule, a signal processing submodule, a Transformer denoising submodule, a distance estimation submodule, and a PPO reinforcement learning positioning submodule. The photodetector submodule receives visible light signals emitted by the signal transmitting module of a mid-altitude reference UAV. The signal processing submodule performs windowing, FFT transformation, and RSS vector extraction on the received signal. The Transformer denoising submodule optimizes the RSS vector for denoising. The distance estimation submodule calculates the distance estimate based on a power function decay model. The PPO reinforcement learning positioning submodule adaptively outputs positioning parameters and calculates the three-dimensional coordinates of the low-altitude UAV using weighted least squares. Low-altitude UAV communication decoding module: Equipped on a low-altitude UAV awaiting positioning, it includes a frequency band separation submodule, a three-level joint decoding submodule, and a fault tolerance verification submodule. The frequency band separation submodule is used to separate the high-frequency positioning signal and the low-frequency communication signal from the visible light signal. The three-level joint decoding submodule is used to sequentially analyze the frequency domain, time-frequency, and instantaneous parameters of the low-frequency communication signal using Fourier transform, wavelet transform, and Hilbert transform. The fault tolerance verification submodule is used to ensure communication reliability and restore communication data through forward error correction codes and cyclic redundancy check. Low- and medium-altitude collaborative control module: used to realize vertical collaboration between low-altitude reference UAVs and low-altitude UAVs to be positioned, network collaboration of multiple low-altitude reference UAVs, and parallel scheduling of navigation and communication functions, including base station scheduling of low-altitude reference UAVs, temporary anchor point management of positioned low-altitude UAVs, and parallel scheduling of positioning and communication functions.

[0028] This application also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement... Figure 1 The methods provided in each step are detailed in the implementation methods provided in the above steps, and will not be repeated here.

[0029] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles (UAVs) based on visible light positioning, characterized in that, Includes the following steps: S1. Positioning and Signal Transmission of the Medium-Altitude Reference UAV: ​​The medium-altitude reference UAV obtains its precise three-dimensional position through image matching positioning. Combined with the three-dimensional position, the absolute spatial coordinates of each LED in the multiple LED arrays carried by the UAV are determined. Frequency division multiplexing technology is used to control the LED arrays carried by the medium-altitude reference UAV to transmit a superimposed signal that simultaneously carries high-frequency positioning signals and low-frequency communication signals downward, forming a visible light signal that can be stably propagated and effectively received within the coverage radius. S2. Low-altitude UAV navigation and positioning calculation: The low-altitude UAV to be positioned receives the visible light signal through a single photodetector on board, and sequentially performs signal windowing, FFT transformation and RSS vector extraction operations. Then, it is optimized by the Transformer deep learning network for noise reduction and distance estimation by the power function decay model. Combined with the PPO reinforcement learning algorithm, the positioning parameters are adaptively output. Finally, the precise three-dimensional coordinates of the UAV are determined by weighted least squares solution. S3. Low-altitude UAV communication decoding: After receiving the visible light signal transmitted in step S1, the low-altitude UAV to be located performs preprocessing, three-level joint decoding and fault-tolerant verification operations in sequence to accurately parse the communication data and realize two-way communication between low- and medium-altitude UAVs for receiving flight control commands, transmitting mission parameters and providing status information feedback.

2. The method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to claim 1, characterized in that, Step S1, which uses frequency division multiplexing technology to control the LED array on the mid-altitude reference UAV to emit a superimposed signal carrying both positioning and communication signals downwards, forming a visible light signal that can propagate stably and be effectively received within its coverage radius, includes: The relative position of each LED in the array is pre-calibrated based on the precise three-dimensional position of the drone, and the absolute spatial coordinates of each LED are determined. Each LED is assigned multiple non-overlapping modulation frequencies to generate a high-frequency positioning signal in the 200-3000Hz band; at the same time, the low-frequency communication signal in the 1-20Hz band is encoded using a dual modulation method of "different frequencies corresponding to flight commands and brightness changes encoding binary data", and then superimposed on the positioning signal of each LED. The LED array is controlled to emit a superimposed signal that simultaneously carries a high-frequency positioning signal of 200-3000Hz and a low-frequency communication signal of 1-20Hz, forming a visible light signal with a coverage radius of 20-50 meters. The positioning signal and communication signal frequency bands are separated and the frequency interval is more than 100 times.

3. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to claim 2, characterized in that, Step S2 specifically includes: S21. The photodetector acquires the time-domain data sequence of the visible light signal at a preset sampling rate, performs windowing processing on the time-domain data sequence, and performs fast Fourier transform frequency domain analysis to extract the received signal strength RSS vector corresponding to each LED modulation frequency. S22. Input the RSS vector into the Transformer deep learning network, correct outliers by learning the spatial correlation between various LED signals, and output the denoised RSS vector. S23. Based on the power function decay model, the distance estimate from the low-altitude UAV to each LED is calculated using the denoised RSS vector. S24. The positioning parameters are adaptively optimized by using the PPO reinforcement learning optimization algorithm. Combining the distance estimate with the known spatial position of the LED, the three-dimensional coordinates of the low-altitude UAV to be positioned are calculated using the weighted least squares algorithm.

4. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to claim 3, characterized in that, Step S21 includes: S211, the photodetector uses a preset sampling rate Continuously acquire time-domain data sequences of visible light signals; S212, for lengths of The time-domain signal is windowed, and the window size is... The number of windows is ), The total length of the time-domain signal; S213, Execute on the data of each window. Point Fast Fourier Transform (FFT) and calculate the one-sided magnitude spectrum; S214. Determine the frequency resolution based on the sampling rate and window size. The calculation formula is: S215. Calculate the corresponding frequency index for the modulation frequency of each LED. The calculation formula is: S216. Frequency index corresponding to the modulation frequency of each LED The peak value is found within ±2 frequency points nearby as the RSS value, and finally formed 3D RSS vector .

5. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to claim 4, characterized in that, The structure of the Transformer deep learning network described in step S22 includes: Input layer: RSS vector , dimension is ; Location encoding layer: for position pos and dimension 2 The formula for calculating the location code is: in, The dimension of the Transformer model; This indicates a position of pos and a dimension of 2. The position encoding result at that time; This indicates that the position is pos and the dimension is The position encoding result at that time; Indicates the index of the current dimension; Transformer encoder layer: Stacked 4 Transformer Encoder layers, each layer includes a multi-head self-attention mechanism, a feedforward neural network, layer normalization and residual connections; Output layer: Outputs the denoised RSS vector. The RSS value affected by noise pollution is corrected to a value that conforms to physical laws.

6. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to claim 5, characterized in that, The expression for the power function decay model in step S23 is: in, For low-altitude drones to the first The distance between LEDs; For the denoised first The received signal strength of each LED; , All parameters were obtained through calibration experiments. The distance calculation involves substituting the denoised RSS values ​​of each LED into the power function attenuation model to obtain the distance. Several distance estimates are used to form a distance vector. .

7. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to claim 6, characterized in that, Step S24 specifically includes: S241. Constructing the state space: at time... Define system state The system state The comprehensive expression, which incorporates information related to current signal characteristics, positioning error, and environmental interference, is as follows: in: The received signal strength of each LED after denoising is output in step S22; The distance estimates from the low-altitude UAV to be located to each LED are calculated in step S23. The standard deviation of the denoised RSS vector; They are time points , The positioning error; SNR is the signal-to-noise ratio of the current visible light signal; S242. Define the action space: the actions output by the PPO policy network. It contains the key parameters of the localization algorithm, expressed as: in: For the first The weighting coefficients of each LED satisfy the constraints. ( =1,…, ); For regularization parameters; To optimize the step size factor; S243. To achieve synergistic optimization of positioning accuracy, trajectory smoothness, and positioning security, a reward function is designed. The expression is: in: This represents the positioning error at the current moment. This is a smoothing penalty term for the positioning trajectory; This is a penalty item for exceeding the limit; , All are preset positive numbers, which are the weight coefficients of the smoothing penalty term and the over-limit penalty term, respectively; S244, PPO Strategy Update: The Actor network, employing an Actor-Critic architecture, outputs localization parameters, while the Critic network evaluates the current system state. The value of this is introduced by the pruning objective function. The policy update range is limited; the policy parameters of the Actor network are updated by maximizing the pruning objective function, and the generalized advantage estimation (GAE) method is combined to stably estimate the advantage function. S245. Based on the positioning parameters output by the PPO policy network after the update in step S244, and combined with the distance estimates obtained in step S23 and the known absolute spatial coordinates of each LED, a weighted residual function is constructed; using the three-dimensional position of the UAV to be positioned as the optimization variable, ... The optimization objective is to minimize the sum of the weighted residual squares corresponding to each LED and the sum of the position regularization term. The nonlinear least squares method is used for iterative solution to obtain the accurate three-dimensional coordinates of the UAV to be located. S246. The three-dimensional coordinate positioning result of the low-altitude UAV obtained in step S245 is directly output to the UAV flight control system to realize the real-time navigation control of the UAV. At the same time, the positioning error between the current positioning result and the actual position is calculated, and the positioning error is fed back to the PPO reinforcement learning network as a feedback signal for the network's online learning, so as to realize the dynamic iteration of positioning parameters and the continuous improvement of positioning accuracy.

8. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to claim 7, characterized in that, In step S244, the target function is pruned. The expression is: in: Importance sampling ratio: This is the clipping function; The dominant function; The weighted residual function expression in step S245 is: in, For the first The weighted residuals corresponding to each LED; For the drone to be located to the first The geometric distance between the LEDs; Distance estimated based on RSS; For the first The weighting coefficient of each LED; The objective function expression is: in, To find the optimal drone location This minimizes the value of the subsequent function. express The sum of squares of the weighted residuals of each LED; Represents a regular term; These are the regularization parameters output by PPO. It's the location of the drone. It's an initial position guess.

9. A method for cooperative navigation and communication of low- and medium-altitude unmanned aerial vehicles based on visible light positioning according to any one of claims 1-8, characterized in that, The three-level decoding in step S3 includes: First-level decoding extracts the low-frequency spectral components of the visible light signal through fast Fourier transform; The second-level decoding uses Morlet wavelets as the basis function to perform continuous wavelet transform, calculates wavelet coefficients, and determines the occurrence time and switching time of significant frequency components through the time-frequency distribution of wavelet coefficients, captures dynamic command sequences, and identifies the amplitude modulation time pattern. Three-level decoding: The instantaneous amplitude, instantaneous frequency, and instantaneous phase of the communication signal are extracted through Hilbert transform, so as to achieve accurate analysis of the dual-modulated signal.

10. A low-to-medium altitude unmanned aerial vehicle (UAV) cooperative navigation and communication system based on visible light positioning, characterized in that, The method for implementing the low- and medium-altitude UAV cooperative navigation and communication method based on visible light positioning as described in any one of claims 1-9 includes: a low-altitude reference UAV signal transmission module, a low-altitude UAV navigation and positioning module, a low-altitude UAV communication decoding module, and a low- and medium-altitude cooperative control module. A mid-altitude reference UAV signal transmission module: mounted on a mid-altitude reference UAV, comprising a positioning submodule, an LED array modulation submodule, and a communication coding submodule; the positioning submodule is used to obtain the precise three-dimensional position of the mid-altitude reference UAV through image matching positioning; the LED array modulation submodule is used to carry multiple LEDs in a preset layout, calibrate the absolute spatial coordinates of each LED according to the three-dimensional position obtained by the positioning submodule, generate a high-frequency positioning signal using frequency division multiplexing technology, and drive the LED array to transmit; the communication coding submodule is used to generate a low-frequency communication signal and perform dual encoding, superimposing it on the positioning signal; Low-altitude UAV navigation and positioning module: Equipped on a low-altitude UAV to be positioned, it includes a photodetector submodule, a signal processing submodule, a Transformer denoising submodule, a distance estimation submodule, and a PPO reinforcement learning positioning submodule. The photodetector submodule receives visible light signals emitted by the signal transmitting module of a mid-altitude reference UAV. The signal processing submodule performs windowing, FFT transformation, and RSS vector extraction on the received signal. The Transformer denoising submodule optimizes the RSS vector for denoising. The distance estimation submodule calculates the distance estimate based on a power function decay model. The PPO reinforcement learning positioning submodule adaptively outputs positioning parameters and calculates the three-dimensional coordinates of the low-altitude UAV using weighted least squares. Low-altitude UAV communication decoding module: Equipped on a low-altitude UAV awaiting positioning, it includes a frequency band separation submodule, a three-level joint decoding submodule, and a fault tolerance verification submodule. The frequency band separation submodule is used to separate the high-frequency positioning signal and the low-frequency communication signal from the visible light signal. The three-level joint decoding submodule is used to sequentially analyze the frequency domain, time-frequency, and instantaneous parameters of the low-frequency communication signal using Fourier transform, wavelet transform, and Hilbert transform. The fault tolerance verification submodule is used to ensure communication reliability and restore communication data through forward error correction codes and cyclic redundancy check. Low- and medium-altitude collaborative control module: used to realize vertical collaboration between low-altitude reference UAVs and low-altitude UAVs to be positioned, network collaboration of multiple low-altitude reference UAVs, and parallel scheduling of navigation and communication functions, including base station scheduling of low-altitude reference UAVs, temporary anchor point management of positioned low-altitude UAVs, and parallel scheduling of positioning and communication functions.

Images