Unmanned aerial vehicle flight control method for preventing wireless interference

Abstract

The invention discloses an unmanned aerial vehicle flight control method for preventing wireless interference, and belongs to the technical field of multi-unmanned aerial vehicle control. The method comprises the following steps: acquiring and preprocessing an SDR signal; performing interference identification on the preprocessed signal and evaluating an interference identification result; a control strategy optimization framework is constructed based on reinforcement learning, and dynamic updating and optimization of flight control parameters are realized; through a multi-mode communication link and a data temporary storage mechanism, real-time monitoring of a task state in an extreme interference scene is ensured. According to the invention, by constructing a multi-sensor fusion autonomous navigation system, high-precision autonomous positioning and dynamic trajectory correction capability are realized; by introducing an intelligent spectrum decision-making mechanism combining software defined radio SDR and machine learning and utilizing fast Fourier transform FFT and reinforcement learning algorithms, full-band interference identification and adaptive frequency hopping communication capabilities are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-UAV control, and in particular relates to a UAV flight control method that is resistant to wireless interference. Background Art

[0002] With the continuous development of drone technology, drones are increasingly being used in military, agriculture, logistics, traffic monitoring and other fields. However, in complex environments, drones may encounter radio interference while flying. This interference can affect communications between the drone and ground control stations or other equipment, preventing it from properly receiving navigation signals (such as GPS / Beidou / RTK) and flight commands. In severe cases, it may cause the drone to lose control or fail to complete its mission.

[0003] Most current technologies rely on satellite navigation systems like GPS and Beidou for drone positioning and navigation. However, GPS and Beidou signals can be affected by factors such as buildings, mountainous areas, and weather. In particular, when encountering radio interference, drones may be unable to complete their missions using traditional navigation methods. Ensuring safe flight and mission completion in such environments remains a pressing issue.

[0004] A Chinese patent application (application number: 201520362380.3) discloses an interference-resistant drone system that improves its adaptability in complex environments through hardware architecture optimization. The drone utilizes a modular design, with the flight control system, image transmission module, and power unit independently packaged and physically shielded to reduce electromagnetic coupling interference between digital and analog circuits, as well as between high- and low-power modules. The communication module integrates 2.4G, 5.8G, and 1.2G multi-band transceivers, and incorporates dynamic frequency hopping technology to isolate the frequencies of control signals and image transmission, reducing co-channel interference. Ground equipment is equipped with a multi-band receiver, forming a reliable, frequency-coordinated communication link with the drone. Furthermore, the frame and landing gear are connected by a rubber gasket with a protruding ring, effectively buffering the effects of vibration on the sensor. This solution, through a combination of multi-band communication, module isolation, and mechanical shock absorption, addresses the sensor accuracy issues associated with conventional drones caused by electromagnetic interference, signal crosstalk, and vibration, significantly improving communication stability and device reliability in complex environments.

[0005] A Chinese patent application (Application No. 201710605267.7) discloses an anti-interference flight control method and device, which utilizes a dual-loop architecture to achieve robust control. The outer loop, based on multi-source positioning data, employs a stabilization-augmented integral backstepping algorithm to compensate for displacement deviations and suppress the impact of external interference such as airflow and noise on the trajectory. The inner loop calculates roll, pitch, and yaw angles in real time, generating attitude control variables using a backstepping robust adaptive algorithm to address attitude instability caused by body vibration and parameter changes. The control device integrates a nonlinear complementary filter module to enhance sensor fusion accuracy and dynamically adjusts parameters through an integral term to reduce overshoot and chatter. This solution forms a full-process anti-interference system from algorithm to device, breaking through the bottleneck of traditional PID control. It significantly enhances the flight stability and control accuracy of drones in complex environments and effectively addresses both deterministic and uncertain interference.

[0006] The patent for an anti-interference drone system has the following shortcomings: 1. It focuses on hardware-level anti-interference design, relying on multi-band communication modules and physical isolation structures, which may increase system complexity and cost, and does not involve the optimization of flight control algorithms;

[0007] 2. For complex and dynamically changing electromagnetic interference scenarios, it is difficult to achieve real-time dynamic suppression of interference through hardware frequency isolation and module shielding alone. The mechanical shock absorption design has limited adaptability to high-frequency vibration or sudden impact, which may affect the long-term monitoring accuracy of the sensor.

[0008] The patented anti-interference flight control method and device have the following shortcomings: 1. It relies on a complex backstepping robust adaptive and stabilization integral strategy, which is computationally intensive and places high demands on the UAV's onboard processor performance, potentially increasing hardware costs and power consumption;

[0009] 2. In the event of strong nonlinear interference or sensor failure, the dynamic adjustment mechanism of the integral term and adaptive parameters may have a response delay, resulting in a decrease in control accuracy, making parameter tuning in actual engineering applications more difficult.

[0010] Therefore, how to improve the accuracy of drone control in a complex electromagnetic environment and reduce costs and power consumption is a technical problem that needs to be solved in this application. Summary of the Invention

[0011] The purpose of the present invention is to provide a UAV flight control method that is resistant to wireless interference, so as to solve the problems raised in the above background technology.

[0012] The object of the present invention is achieved as follows: a method for controlling the flight of an unmanned aerial vehicle (UAV) with anti-wireless interference, characterized in that the method comprises the following steps:

[0013] Step S1: collecting and preprocessing the SDR signal;

[0014] Step S2: performing interference identification on the pre-processed signal and evaluating the interference identification result;

[0015] Step S3: Build a control strategy optimization framework based on reinforcement learning to achieve dynamic update and optimization of flight control parameters;

[0016] Step S4: Ensure real-time monitoring of task status in extreme interference scenarios through multimodal communication links and data temporary storage mechanisms.

[0017] Preferably, in step S1, the SDR signal is collected and preprocessed, specifically:

[0018] Step S1-1: bandpass filtering the obtained SDR signal;

[0019] Step S1-2: Perform FFT transformation using overlapping segment processing to generate a time-frequency spectrum. Specifically:

[0020] First, the signal is divided into multiple fixed-length and overlapping data frames, and a window function is applied to each data frame for point-by-point weighted processing;

[0021] Perform fast Fourier transform on each frame of weighted data to transform it from time domain to frequency domain, and obtain the complex frequency X containing amplitude and phase information within the time period. m [k], where m is the index of the time frame and k is the index of the frequency;

[0022] The power spectrum S[m,k]=|X is obtained by calculating the square of the complex spectrum amplitude m [k]| 2 , all calculated power values ​​S[m,k] are arranged in their corresponding time sequence to form a two-dimensional time-frequency spectrum diagram;

[0023] Step S1-3: Use an adaptive threshold detection algorithm to identify abnormal spectrum peaks and energy distribution.

[0024] Preferably, in step S1-3, an adaptive threshold detection algorithm is used to identify abnormal spectrum peaks and energy distribution, specifically:

[0025] For any point (m, k) to be measured in the generated time-spectrum diagram, the local background noise estimation value P of the point is obtained by calculating the power average of other points in its surrounding neighborhood. noise [m,k];

[0026] Based on the noise estimate, the dynamic detection threshold T[m,k] is calculated by the following relationship:

[0027] T[m,k]=γ·P noise[m,k]+δ;

[0028] Where γ is a preset gain factor greater than 1, which is used to ensure that the threshold is always above the noise baseline, and δ is the protection offset;

[0029] The power S[m,k] of the point to be tested is compared with the calculated dynamic threshold T[m,k]. If the former is greater than the latter, the point is identified and marked as an abnormal spectrum point, indicating the presence of potential radio interference.

[0030] Preferably, in step S2, interference identification is performed on the pre-processed signal and the interference identification result is evaluated, specifically:

[0031] Step S2-1: Extract time domain features, frequency domain features, and statistical features in sequence through a multi-level feature extraction network, specifically:

[0032] According to the position of the interference area on the time-frequency spectrum, locate its coordinate interval on the time axis, intercept the corresponding original I / Q signal segment and two-dimensional time-frequency spectrum, and uniformly feed them into the multi-level feature extraction network;

[0033] For one-dimensional time-domain signal sequences, the multi-level feature extraction network uses a one-dimensional convolution kernel to perform sliding convolution operations along the time axis to capture the instantaneous features of pulses, edges, and short-term energy changes in the signal;

[0034] For the two-dimensional time-frequency spectrum, the multi-level feature extraction network uses a two-dimensional convolution kernel to perform sliding convolution operations on the graph to identify the static texture features of the spectrum;

[0035] After each convolution operation, nonlinearity is added through activation functions and downsampling is performed through pooling layers to reduce the data dimension while retaining key features;

[0036] The time domain features and frequency domain features output by the two parallel convolution processing modules are spliced ​​in the fusion layer built into the multi-level feature extraction network to integrate them into a fused feature vector sequence containing joint time and frequency information;

[0037] The fused feature vector sequence is fed into the recurrent neural network layer in the multi-level feature extraction network in chronological order. The recurrent neural network layer uses LSTM units to learn and extract the dynamic laws of the interference signal evolving over time.

[0038] The multi-level feature extraction network outputs the hidden state of the recurrent neural network module at the last time step. This state is the final composite feature vector containing time domain, frequency domain and dynamic information, and is transmitted to the subsequent classifier for processing;

[0039] Step S2-2: Send the feature vector to the hierarchical classifier to determine the presence of interference and subdivide the interference type and severity;

[0040] The hierarchical classifier adopts a multi-layer perceptron network, which performs classification tasks based on the input feature vector and directly outputs a clear interference state label;

[0041] Step S2-3: Based on the sequential state machine model, the multi-frame judgment results are integrated to eliminate instantaneous misjudgments and finally output a stable interference state assessment result.

[0042] Preferably, in step S2-3, based on the sequential state machine model, the multi-frame judgment results are integrated to eliminate instantaneous misjudgments, and finally a stable interference state evaluation result is output, specifically:

[0043] At each time frame, the sequential state machine model receives the instantaneous decision result of the hierarchical classifier and compares it with the current confirmation state maintained inside the sequential state machine model;

[0044] If the instantaneous judgment result is the same as the current confirmation status, all confirmation counters that are counting will be reset and the current confirmation status will remain unchanged;

[0045] If the instantaneous judgment result is different from the current confirmation status, the following judgment and operation are performed:

[0046] First, check whether there is a confirmation counter counting for the new state corresponding to the instantaneous judgment result; if not, initialize a confirmation counter for the new state and set its value to 1; at the same time, reset the confirmation counters of all other states;

[0047] If the corresponding confirmation counter already exists, its count value is increased by one;

[0048] Then, determine whether the value of the counter has reached the preset continuous occurrence threshold N:

[0049]

[0050] Among them, P target is the target false alarm probability of the timing state machine, P fa is the false alarm probability of a single frame with interference;

[0051] If the count value has reached the threshold value N, the current confirmation state is officially updated to the new interference state, and the counter is reset; if the count value has not reached the threshold value N, the current confirmation state remains unchanged.

[0052] Preferably, in step S3, a control strategy optimization framework is constructed based on reinforcement learning to achieve dynamic updating and optimization of flight control parameters, specifically:

[0053] Step S3-1: The control strategy optimization framework adopts a hierarchical control architecture, which divides the flight control into trajectory planning layer, dynamic control layer and bottom execution layer;

[0054] Trajectory Planning Layer: As the highest decision-making layer, the trajectory planning layer is responsible for long-term global path generation and macro-decision-making, specifically:

[0055] First, a dynamic cost map is constructed that represents the flight environment as a three-dimensional grid. The cost value of each grid in the dynamic cost map consists of two parts: a static base cost reflecting the terrain or fixed no-fly zones; and a dynamically changing interference cost. The size of this cost value is directly related to the interference intensity.

[0056] Subsequently, this method calls the A-Star path search algorithm, using the current position of the UAV as the starting point and the mission target as the end point, to search on the dynamic cost map;

[0057] After the search is completed, the resulting path is converted into a smooth and executable flight trajectory sequence consisting of spatial coordinate points and sent to the dynamic control layer;

[0058] Dynamic Control Layer: This layer is responsible for high-frequency real-time flight status tracking. Its core task is to convert the target waypoints and speed instructions provided by the trajectory planning layer into low-level control instructions that the drone can directly execute.

[0059] The workflow begins with error calculation. The dynamic control layer continuously compares the desired state of the target position and velocity from the trajectory planning layer with the actual state from the onboard sensors, thereby generating a real-time position and velocity error vector.

[0060] Subsequently, the internal control law of the dynamic control layer calculates the three-dimensional vector of the resultant force and torque that must be applied to correct the error and track the desired trajectory based on this error vector;

[0061] Finally, the system decomposes this abstract force and torque vector into specific, standardized control instructions;

[0062] Execution layer: The bottom execution layer receives attitude and thrust commands from the dynamic control layer and directly converts them into pulse width modulation (PWM) signals that control the speed of each motor, thereby driving the drone to complete actual physical actions;

[0063] Step S3-2: Design an adaptive control gain adjustment mechanism to dynamically adjust control parameters according to flight status and environmental interference level;

[0064] Step S3-3: Use a multi-level security mechanism to perform global path planning.

[0065] Preferably, in step S3-3, a multi-level security mechanism is used to perform global path planning, specifically:

[0066] In the trajectory planning layer, three progressive safety planning modes are set to deal with different levels of radio interference and navigation signal loss risks;

[0067] Level 1 Normal Mode: When there is no interference or low-level interference and the main navigation signal is normal, the path planner executes the mission according to the preset route or the optimal economic route;

[0068] Level 2 Avoidance Mode: This mode is activated in the event of medium to high interference intensity. Based on the possible location of the interference source or the affected area, the interference source is marked as a high-cost "virtual obstacle" in the path planning algorithm's cost map, and a safe, cost-effective route is automatically replanned to bypass the area.

[0069] Level 3 self-preservation mode: This mode is activated when there is extreme interference and the primary navigation signal is completely lost. The system immediately terminates the current mission and switches to the backup autonomous navigation mode to execute the preset emergency plan.

[0070] Preferably, in step S4, a multimodal communication link and a data temporary storage mechanism are used to ensure real-time monitoring of the task status in extreme interference scenarios, specifically:

[0071] Step S4-1: compress and encode the original sensor data and store them in a hierarchical manner according to their importance, specifically:

[0072] Level 1 is core flight safety data including attitude, position, speed and control mode;

[0073] Level 2 is critical data generated by mission payloads; Level 3 is routine equipment telemetry data;

[0074] A unified, efficient, lossless compression algorithm is used to compress all classified data. The compressed data is then temporarily stored in an onboard circular buffer. The buffer adopts a hierarchical storage strategy, allocating protected, last-overwritten storage areas for primary data. This ensures that even when communications are interrupted and storage space is limited, the most critical flight records of the drone are preserved with the highest priority and in the most complete manner.

[0075] Step S4-2: Adopt an adaptive channel selection mechanism to continuously evaluate the quality of each communication channel and prioritize the most reliable channel for data transmission. Specifically:

[0076] The adaptive channel selection mechanism starts a timer that is set to trigger every 100 milliseconds to drive the loop execution of the entire selection process;

[0077] At the beginning of each cycle, all available communication links are queried in parallel, and the original values ​​of the three key performance indicators (KPIs) of each link at the current moment, namely, the received signal strength indicator RSSI, the signal-to-noise ratio SNR, and the packet loss rate PLR, are obtained;

[0078] For each link i, the collected original value RSSI is normalized using the “min-max normalization” formula i and SNR i Convert to a standardized value in the [0,1] interval;

[0079] The comprehensive quality score Q of each link is calculated through the preset weighted formula i :

[0080] Q i =w R norm(RSSI i )+w S norm(SNR i )-w P PLR i ;

[0081] Among all the link quality scores Q, the link with the highest score is determined. When its score is greater than the current link score, the link switching procedure is immediately executed to switch the main communication task to the link with the highest score.

[0082] Step S4-3: Execute a state smooth transition algorithm to gradually fuse the state estimation in the emergency mode with the recovered satellite positioning data.

[0083] Preferably, in step S4-3, a state smooth transition algorithm is executed to gradually fuse the state estimation in the emergency mode with the recovered satellite positioning data. The state smooth transition algorithm adopts a linear weighted fusion algorithm based on a time window to achieve smooth transition, which is specifically implemented as follows:

[0084] After confirming that the GNSS signal has stabilized, the process is triggered and a transition time window with a fixed duration of T is started. The start time t0 is recorded. In each control cycle within the time window, the following operations are performed:

[0085] First, read the latest autonomous navigation position P VIO (t), and the current GNSS position P GNSS (t);

[0086] According to the current time t, calculate a weight factor that increases linearly from 0 to 1

[0087] By weighting, the output position command P(t) that should be used in the current cycle is calculated:

[0088] P(t)=(1-β(t))·P VIO (t)+β(t)·P GNSS (t);

[0089] The output position command P(t) calculated in the previous step is used as the final position target of the current control cycle and sent to the underlying flight controller for execution. When time t reaches or exceeds t0+T, the transition time window ends.

[0090] The satellite position P GNSS (t) directly outputs the position command and continues to use this mode until the next navigation source switch occurs; the entire smooth transition process is now complete;

[0091] When the system confirms that the GNSS signal has returned to stability, the linear weighted fusion algorithm is started within a preset fixed time window;

[0092] At the moment of startup, the linear weighted fusion algorithm first calculates and locks the initial deviation between the current airborne autonomous navigation position and the true position;

[0093] Throughout the time window, a linear weighted fusion algorithm continuously generates a blended position and provides it to the flight control system. This blended position is calculated through a dynamically changing weighted averaging process: at the start of the time window, the weight is completely biased towards the onboard autonomous navigation position, with the VIO position weighted at 100% and the GNSS position weighted at zero. As time passes, the weight of the onboard autonomous navigation position decreases linearly and smoothly from 100% to zero, while the weight of the GNSS position increases linearly and smoothly from zero to 100% simultaneously.

[0094] When the time window ends, the weight of the position estimation has been completely transferred to GNSS, thus completing a continuous and smooth state transition and ensuring the stability of flight control.

[0095] Compared with the prior art, the present invention has the following improvements and advantages:

[0096] 1. By building an autonomous navigation system with multi-sensor fusion, high-precision autonomous positioning and dynamic trajectory correction capabilities are achieved. By introducing an intelligent spectrum decision-making mechanism that combines software-defined radio (SDR) with machine learning, and utilizing fast Fourier transform (FFT) and reinforcement learning algorithms, full-band interference recognition and adaptive frequency hopping communication capabilities are achieved.

[0097] 2. By integrating visual and radar data to build a real-time environmental map, and using reinforcement learning control strategies, it can achieve omnidirectional three-dimensional environmental perception and high-speed dynamic obstacle avoidance capabilities; through multimodal communication links and data temporary storage mechanisms, using LTE / 5G encrypted transmission protocols and flash storage technology, it can achieve anti-interference communication links and real-time return of mission data. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 Schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0099] The present invention is further summarized below with reference to the accompanying drawings.

[0100] like Figure 1 As shown, a UAV flight control method for preventing wireless interference includes the following steps:

[0101] Step S1: collecting and preprocessing the SDR signal;

[0102] Step S1-1: bandpass filtering the obtained SDR signal;

[0103] The digital signal stream collected by the SDR is fed into a digital bandpass filter. The passband range of the filter is pre-set according to the center frequency and bandwidth of the UAV communication link. Its function is to allow only the signal components within the specific frequency band to pass through and attenuate all out-of-band signals and noise. After this step, a signal sequence that is concentrated in the target frequency band and has undergone preliminary purification is obtained. The passband range of the filter is pre-set according to the center frequency and bandwidth of the UAV communication link. Its function is to allow only the signal components within the specific frequency band to pass through and attenuate all out-of-band signals and noise. After this step, a signal sequence that is concentrated in the target frequency band and has undergone preliminary purification is obtained.

[0104] Step S1-2: Perform FFT transformation using overlapping segment processing to generate a time-frequency spectrum. Specifically:

[0105] First, the signal is divided into multiple fixed-length and overlapping data frames, and a window function is applied to each data frame for point-by-point weighted processing;

[0106] Perform fast Fourier transform on each frame of weighted data to transform it from time domain to frequency domain, and obtain the complex frequency X containing amplitude and phase information within the time period. m [k], where m is the index of the time frame and k is the index of the frequency;

[0107] The power spectrum S[m,k]=|X is obtained by calculating the square of the complex spectrum amplitude m [k]| 2 , all calculated power values ​​S[m,k] are arranged in their corresponding time sequence to form a two-dimensional time-frequency spectrum diagram;

[0108] Step S1-3: Use an adaptive threshold detection algorithm to identify abnormal spectrum peaks and energy distribution, specifically:

[0109] For any point (m, k) to be measured in the generated time-spectrum diagram, the local background noise estimation value P of the point is obtained by calculating the power average of other points in its surrounding neighborhood. noise [m,k];

[0110] Based on the noise estimate, the dynamic detection threshold T[m,k] is calculated by the following relationship:

[0111] T[m,k]=γ·P noise [m,k]+δ;

[0112] Where γ is a preset gain factor greater than 1, which is used to ensure that the threshold is always above the noise baseline, and δ is the protection offset;

[0113] The power S[m,k] of the point to be tested is compared with the calculated dynamic threshold T[m,k]. If the former is greater than the latter, the point is identified and marked as an abnormal spectrum point, indicating the presence of potential radio interference.

[0114] Step S2: performing interference identification on the pre-processed signal and evaluating the interference identification result;

[0115] Step S2-1: extracting time domain features, frequency domain features and statistical features in sequence through a multi-level feature extraction network;

[0116] According to the position of the interference area on the time-frequency spectrum, locate its coordinate interval on the time axis, intercept the corresponding original I / Q signal segment and two-dimensional time-frequency spectrum, and uniformly feed them into the multi-level feature extraction network;

[0117] For one-dimensional time-domain signal sequences, the multi-level feature extraction network uses a one-dimensional convolution kernel to perform sliding convolution operations along the time axis to capture the instantaneous features of pulses, edges, and short-term energy changes in the signal;

[0118] For the two-dimensional time-frequency spectrum, the multi-level feature extraction network uses a two-dimensional convolution kernel to perform sliding convolution operations on the graph to identify the static texture features of the spectrum;

[0119] After each convolution operation, nonlinearity is added through activation functions and downsampling is performed through pooling layers to reduce the data dimension while retaining key features;

[0120] The time domain features and frequency domain features output by the two parallel convolution processing modules are spliced ​​in the fusion layer built into the multi-level feature extraction network to integrate them into a fused feature vector sequence containing joint time and frequency information;

[0121] The fused feature vector sequence is fed into the recurrent neural network layer in the multi-level feature extraction network in chronological order. The recurrent neural network layer uses LSTM units to learn and extract the dynamic laws of the interference signal evolving over time.

[0122] The multi-level feature extraction network outputs the hidden state of the recurrent neural network module at the last time step. This state is the final composite feature vector containing time domain, frequency domain and dynamic information, and is transmitted to the subsequent classifier for processing;

[0123] Step S2-2: Send the feature vector to the hierarchical classifier to determine the presence of interference and subdivide the interference type and severity;

[0124] The hierarchical classifier uses a multi-layer perceptron network, which performs classification tasks based on the input feature vector and directly outputs a clear interference state label;

[0125] The fused feature vector generated in step S2-1 is input into a pre-trained multi-layer perceptron network. The multi-layer perceptron network performs a unique classification task based on the input feature vector and directly outputs a clear interference status label. This label is selected from a predefined set, such as no interference, low-intensity single-tone interference, high-intensity single-tone interference, broadband suppression interference, frequency-hopping interference, etc.

[0126] Step S2-3: Based on the sequential state machine model, the multi-frame judgment results are integrated to eliminate instantaneous misjudgments and finally output a stable interference state assessment result, specifically:

[0127] This step uses a sequential state machine model to filter out isolated instantaneous determination results output by step S2-2. This sequential state machine model achieves this by internally maintaining a current confirmation state as a stable output and establishing a confirmation counter for each potential new state. It takes the instantaneous determination result from S2-2 as input and decides whether to update the current confirmation state based on a preset continuous occurrence threshold N, thereby outputting a stable and reliable interference state assessment result that has undergone consistency verification over time.

[0128] At each time frame, the sequential state machine model receives the instantaneous decision result of the hierarchical classifier and compares it with the current confirmation state maintained inside the sequential state machine model;

[0129] If the instantaneous judgment result is the same as the current confirmation status, all confirmation counters that are counting will be reset and the current confirmation status will remain unchanged;

[0130] If the instantaneous judgment result is different from the current confirmation status, the following judgment and operation are performed:

[0131] First, check whether there is a confirmation counter counting for the new state corresponding to the instantaneous judgment result; if not, initialize a confirmation counter for the new state and set its value to 1; at the same time, reset the confirmation counters of all other states;

[0132] If the corresponding confirmation counter already exists, its count value is increased by one;

[0133] Then, determine whether the value of the counter has reached the preset continuous occurrence threshold N:

[0134]

[0135] Among them, P target is the target false alarm probability of the timing state machine, P fa is the false alarm probability of a single frame with interference;

[0136] If the count value has reached the threshold value N, the current confirmation state is officially updated to the new interference state, and the counter is reset; if the count value has not reached the threshold value N, the current confirmation state remains unchanged.

[0137] Step S3: Build a control strategy optimization framework based on reinforcement learning to achieve dynamic update and optimization of flight control parameters;

[0138] Step S3-1: The control strategy optimization framework adopts a hierarchical control architecture, which divides the flight control into trajectory planning layer, dynamic control layer and bottom execution layer;

[0139] Trajectory Planning Layer: As the highest decision-making layer, the trajectory planning layer is responsible for long-term global path generation and macro-decision-making; the trajectory planning layer receives high-level mission objectives set by the ground station and continuously integrates quantitative interference assessment results.

[0140] First, a dynamic cost map is constructed that represents the flight environment as a three-dimensional grid. The cost value of each grid in the dynamic cost map consists of two parts: a static base cost reflecting the terrain or fixed no-fly zones; and a dynamically changing interference cost. The size of this cost value is directly related to the interference intensity.

[0141] Subsequently, this method calls the A-Star path search algorithm, starting with the UAV's current location and the mission target as the end point, and searches on the dynamic cost map. The goal of the A-Star algorithm is to find a path from the start point to the end point such that the cumulative cost of all grids passed on the path is minimized. Since the cost value of the interference area is significantly increased, the algorithm will naturally plan a path to bypass the strong interference area during the search process.

[0142] After the search is completed, the final generated path is converted into a smooth and executable flight trajectory sequence consisting of spatial coordinate points and sent to the dynamic control layer.

[0143] Dynamic Control Layer: The dynamic control layer is responsible for high-frequency real-time flight status tracking. Its core task is to convert the target waypoints and speed instructions provided by the trajectory planning layer into low-level control instructions that the drone can directly execute. The specific workflow begins with error calculation. The dynamic control layer continuously compares the expected state of the target position and speed from the trajectory planning layer with the actual state from the onboard sensors, thereby generating a real-time position and velocity error vector.

[0144] Subsequently, the internal control law of the dynamic control layer calculates the three-dimensional vector of the resultant force and torque that must be applied to correct the error and track the desired trajectory based on this error vector.

[0145] Finally, the system decomposes this abstract force and torque vector into specific, standardized control instructions, which are then passed to the underlying execution layer.

[0146] Bottom execution layer: The bottom execution layer receives attitude and thrust commands from the dynamic control layer and directly converts them into pulse width modulation (PWM) signals that control the speed of each motor, thereby driving the drone to complete actual physical movements.

[0147] Step S3-2: Design an adaptive control gain adjustment mechanism to dynamically adjust the control parameters according to the flight status and environmental interference level. Specifically:

[0148] Optimize the PID controller gains in the attitude control loop in real time using a reinforcement learning agent:

[0149] State space: The observed state of the RL agent consists of two parts: one is the interference state label, which describes the severity of the external environment; the other is the flight state error of the drone itself, including the deviation between the desired and actual postures, and the rate of change of this deviation.

[0150] Action Space: The output action of the RL agent is a set of adjustment coefficients applied to the three gains (P, I, D) of the PID controller; these coefficients do not directly replace the original gains, but rather fine-tune them to achieve smooth and stable parameter transitions.

[0151] Reward function: This function is designed to reward desired behavior and penalize undesirable behavior. Specifically, the ability to quickly and accurately track the desired posture without overshoot or oscillation will receive a positive reward. Conversely, large tracking errors or unstable flight postures will result in negative rewards.

[0152] Policy Optimization: Through continuous online or offline training, the RL agent learns an optimal policy; this policy can automatically output a set of optimal PID gain adjustment coefficients based on the current interference environment and flight state, allowing the drone to maintain optimal stability and responsiveness under different interference conditions.

[0153] Step S3-3: Use a multi-level security mechanism to perform global path planning, specifically:

[0154] In the trajectory planning layer, three progressive safety planning modes are set to deal with different levels of radio interference and navigation signal loss risks;

[0155] Level 1 Normal Mode: When there is no interference or low-level interference and the main navigation signal is normal, the path planner executes the mission according to the preset route or the optimal economic route;

[0156] Level 2 Avoidance Mode: This mode is activated in the event of medium to high interference intensity. Based on the possible location of the interference source or the affected area, the interference source is marked as a high-cost "virtual obstacle" in the path planning algorithm's cost map, and a safe, cost-effective route is automatically replanned to bypass the area.

[0157] Level 3 self-preservation mode: This mode is activated when there is extreme interference and the primary navigation signal is completely lost. The system immediately terminates the current mission and switches to the backup autonomous navigation mode to execute the preset emergency plan.

[0158] Step S4: Ensure real-time monitoring of mission status in extreme interference scenarios through multimodal communication links and data temporary storage mechanisms;

[0159] Step S4-1: compress and encode the raw sensor data and store them in a hierarchical manner according to their importance. This step is continuously executed on the airborne end to pre-process the data for subsequent communication or temporary storage. Specifically,

[0160] First, all system and mission data to be sent or recorded are divided into three levels according to their importance: Level 1 is core flight safety data including attitude, position, velocity, and control mode; Level 2 is critical data generated by the mission payload; and Level 3 is routine equipment telemetry data.

[0161] Subsequently, in order to effectively reduce the data volume without significantly increasing the computing burden, the system adopts a unified, efficient, lossless compression algorithm, such as the Lempel-Ziv algorithm, to compress all graded data; the compressed data is sent to an onboard circular buffer for temporary storage; the buffer adopts a hierarchical storage strategy, that is, a protected and last-overwritten storage area is allocated for the first-level data, thereby ensuring that when communication is interrupted and storage space is limited, the most core flight process records of the drone are retained with the highest priority and in the most complete way.

[0162] Step S4-2: Adopt an adaptive channel selection mechanism to continuously evaluate the quality of each communication channel and prioritize the most reliable channel for data transmission. Specifically:

[0163] The adaptive channel selection mechanism starts a timer that is set to trigger every 100 milliseconds to drive the loop execution of the entire selection process;

[0164] At the beginning of each cycle, all available communication links are queried in parallel, and the original values ​​of the three key performance indicators (KPIs) of each link at the current moment, namely, the received signal strength indicator RSSI, the signal-to-noise ratio SNR, and the packet loss rate PLR, are obtained;

[0165] For each link i, the collected original value RSSI is normalized using the “min-max normalization” formula i and SNR i Convert to a standardized value in the [0,1] interval;

[0166] The comprehensive quality score Q of each link is calculated through the preset weighted formula i :

[0167] Q i =w R norm(RSSI i )+w S norm(SNR i )-w P PLR i ;

[0168] Among all the link quality scores Q, the link with the highest score is determined. When its score is greater than the current link score, the link switching procedure is immediately executed to switch the main communication task to the link with the highest score.

[0169] Utilize the multimodal communication links integrated on the drone to build a redundant communication system;

[0170] By continuously monitoring key indicators such as received signal strength, signal-to-noise ratio, and packet loss rate across each link, a real-time, comprehensive quality score is calculated for each link. When transmitting data, the system automatically selects the channel with the highest quality score. This allows it to seamlessly switch data streams to more reliable backup channels if interference with the primary channel occurs, ensuring continuous communication.

[0171] Step S4-3: Execute the state smooth transition algorithm to gradually fuse the state estimation in the emergency mode with the recovered satellite positioning data, specifically:

[0172] After confirming that the GNSS signal has stabilized, the process is triggered and a transition time window with a fixed duration of T is started. The start time t0 is recorded. In each control cycle within the time window, the following operations are performed:

[0173] First, read the latest autonomous navigation position P VIO (t), and the current GNSS position P GNSS (t);

[0174] According to the current time t, calculate a weight factor that increases linearly from 0 to 1

[0175] By weighting, the output position command P(t) that should be used in the current cycle is calculated:

[0176] P(t)=(1-β(t))·P VIO (t)+β(t)·P GNSS (t);

[0177] The output position command P(t) calculated in the previous step is used as the final position target of the current control cycle and sent to the underlying flight controller for execution. When time t reaches or exceeds t0+T, the transition time window ends.

[0178] The satellite position P GNSS (t) is directly used as the output position instruction, and this mode is continuously used until the next navigation source is switched; the entire smooth transition process is now complete.

[0179] This step aims to address a key safety issue faced by drones when switching from emergency navigation mode, which relies entirely on onboard sensors (such as visual-inertial odometry (VIO)), back to regular navigation mode based on the Global Navigation Satellite System (GNSS).

[0180] When the system confirms that the GNSS signal has returned to stability, the linear weighted fusion algorithm is started within a preset fixed time window;

[0181] At the moment of startup, the linear weighted fusion algorithm first calculates and locks the initial deviation between the current airborne autonomous navigation position and the true position;

[0182] Throughout the time window, a linear weighted fusion algorithm continuously generates a blended position and provides it to the flight control system. This blended position is calculated through a dynamically changing weighted averaging process: at the start of the time window, the weight is completely biased towards the onboard autonomous navigation position, with the VIO position weighted at 100% and the GNSS position weighted at zero. As time passes, the weight of the onboard autonomous navigation position decreases linearly and smoothly from 100% to zero, while the weight of the GNSS position increases linearly and smoothly from zero to 100% simultaneously.

[0183] When the time window ends, the weight of the position estimation has been completely transferred to GNSS, thus completing a continuous and smooth state transition and ensuring the stability of flight control.

[0184] Due to the inherent cumulative drift errors of autonomous navigation methods like VIO, their calculated positions often deviate significantly from the recently recovered GNSS position. Directly switching state estimation from VIO to GNSS can cause a sudden jump in the drone's position estimate, potentially triggering the flight control system to issue drastic and unintended control commands, posing a serious threat to flight safety. Once the system detects that the satellite navigation signal has been restored and stabilized, it does not immediately use its position to correct the flight state. At the beginning of the transition process, the system continues to use the position from the onboard autonomous navigation. Subsequently, within a time window, the system's confidence in the satellite positioning data increases continuously and smoothly from zero, while its confidence in the onboard autonomous navigation results decreases accordingly and synchronously. At the end of this time window, confidence is fully transferred to the satellite positioning data, completing the system's position estimate transition to satellite-dominated mode. This ensures an absolutely safe mode transition and continuous and stable flight.

[0185] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A UAV flight control method for preventing wireless interference, characterized by: The method comprises the following steps: Step S1: collecting and preprocessing the SDR signal; Step S2: performing interference identification on the pre-processed signal and evaluating the interference identification result; Step S3: Build a control strategy optimization framework based on reinforcement learning to achieve dynamic update and optimization of flight control parameters; Step S4: Ensure real-time monitoring of task status in extreme interference scenarios through multimodal communication links and data temporary storage mechanisms.

2. The method for controlling a UAV flight with wireless interference prevention according to claim 1, characterized in that: In step S1, the SDR signal is collected and pre-processed, specifically: Step S1-1: bandpass filtering the obtained SDR signal; Step S1-2: Perform FFT transformation using overlapping segment processing to generate a time-frequency spectrum. Specifically: First, the signal is divided into multiple fixed-length and overlapping data frames, and a window function is applied to each data frame for point-by-point weighted processing; Perform fast Fourier transform on each frame of weighted data to transform it from time domain to frequency domain, and obtain the complex frequency X containing amplitude and phase information within the time period. m [k], where m is the index of the time frame and k is the index of the frequency; The power spectrum S[m,k]=|X is obtained by calculating the square of the complex spectrum amplitude m [k]| 2 , all calculated power values ​​S[m,k] are arranged in their corresponding time sequence to form a two-dimensional time-frequency spectrum diagram; Step S1-3: Use an adaptive threshold detection algorithm to identify abnormal spectrum peaks and energy distribution.

3. The method for controlling a UAV flight with wireless interference prevention according to claim 2, characterized in that: In step S1-3, an adaptive threshold detection algorithm is used to identify abnormal spectrum peaks and energy distribution, specifically: For any point (m, k) to be measured in the generated time-spectrum diagram, the local background noise estimation value P of the point is obtained by calculating the power average of other points in its surrounding neighborhood. noise [m,k]; Based on the noise estimate, the dynamic detection threshold T[m,k] is calculated by the following relationship: T[m,k]=γ·P noise [m,k]+δ; Where γ is a preset gain factor greater than 1, which is used to ensure that the threshold is always above the noise baseline, and δ is the protection offset; The power S[m,k] of the point to be tested is compared with the calculated dynamic threshold T[m,k]. If the former is greater than the latter, the point is identified and marked as an abnormal spectrum point, indicating the presence of potential radio interference.

4. The method for controlling a UAV flight with wireless interference prevention according to claim 1, wherein: In step S2, interference identification is performed on the pre-processed signal and the interference identification result is evaluated, specifically: Step S2-1: Extract time domain features, frequency domain features, and statistical features in sequence through a multi-level feature extraction network, specifically: According to the position of the interference area on the time-frequency spectrum, locate its coordinate interval on the time axis, intercept the corresponding original I / Q signal segment and two-dimensional time-frequency spectrum, and uniformly feed them into the multi-level feature extraction network; For one-dimensional time-domain signal sequences, the multi-level feature extraction network uses a one-dimensional convolution kernel to perform sliding convolution operations along the time axis to capture the instantaneous features of pulses, edges, and short-term energy changes in the signal; For the two-dimensional time-frequency spectrum, the multi-level feature extraction network uses a two-dimensional convolution kernel to perform sliding convolution operations on the graph to identify the static texture features of the spectrum; After each convolution operation, nonlinearity is added through activation functions and downsampling is performed through pooling layers to reduce the data dimension while retaining key features; The time domain features and frequency domain features output by the two parallel convolution processing modules are spliced ​​in the fusion layer built into the multi-level feature extraction network to integrate them into a fused feature vector sequence containing joint time and frequency information; The fused feature vector sequence is fed into the recurrent neural network layer in the multi-level feature extraction network in chronological order. The recurrent neural network layer uses LSTM units to learn and extract the dynamic laws of the interference signal evolving over time. The multi-level feature extraction network outputs the hidden state of the recurrent neural network module at the last time step. This state is the final composite feature vector containing time domain, frequency domain and dynamic information, and is transmitted to the subsequent classifier for processing; Step S2-2: Send the feature vector to the hierarchical classifier to determine the presence of interference and subdivide the interference type and severity; The hierarchical classifier adopts a multi-layer perceptron network, which performs classification tasks based on the input feature vector and directly outputs a clear interference state label; Step S2-3: Based on the sequential state machine model, the multi-frame judgment results are integrated to eliminate instantaneous misjudgments and finally output a stable interference state assessment result.

5. The method for controlling a UAV flight with wireless interference prevention according to claim 4, characterized in that: In step S2-3, based on the sequential state machine model, the multi-frame judgment results are integrated to eliminate instantaneous misjudgments and finally output a stable interference state assessment result, specifically: At each time frame, the sequential state machine model receives the instantaneous decision result of the hierarchical classifier and compares it with the current confirmation state maintained inside the sequential state machine model; If the instantaneous judgment result is the same as the current confirmation status, all confirmation counters that are counting will be reset and the current confirmation status will remain unchanged; If the instantaneous judgment result is different from the current confirmation status, the following judgment and operation are performed: First, check whether there is a confirmation counter counting for the new state corresponding to the instantaneous judgment result; if not, initialize a confirmation counter for the new state and set its value to 1; at the same time, reset the confirmation counters of all other states; If the corresponding confirmation counter already exists, its count value is increased by one; Then, determine whether the value of the counter has reached the preset continuous occurrence threshold N: Among them, P target is the target false alarm probability of the timing state machine, P fa is the false alarm probability of a single frame with interference; If the count value has reached the threshold value N, the current confirmation state is officially updated to the new interference state, and the counter is reset; if the count value has not reached the threshold value N, the current confirmation state remains unchanged.

6. The method for controlling a UAV flight with wireless interference prevention according to claim 1, characterized in that: In step S3, a control strategy optimization framework is constructed based on reinforcement learning to achieve dynamic updating and optimization of flight control parameters, specifically: Step S3-1: The control strategy optimization framework adopts a hierarchical control architecture, which divides the flight control into trajectory planning layer, dynamic control layer and bottom execution layer; Trajectory Planning Layer: As the highest decision-making layer, the trajectory planning layer is responsible for long-term global path generation and macro-decision-making, specifically: First, a dynamic cost map is constructed that represents the flight environment as a three-dimensional grid. The cost value of each grid in the dynamic cost map consists of two parts: a static base cost reflecting the terrain or fixed no-fly zones; and a dynamically changing interference cost. The size of this cost value is directly related to the interference intensity. Subsequently, this method calls the A-Star path search algorithm, using the current position of the UAV as the starting point and the mission target as the end point, to search on the dynamic cost map; After the search is completed, the resulting path is converted into a smooth and executable flight trajectory sequence consisting of spatial coordinate points and sent to the dynamic control layer; Dynamic Control Layer: This layer is responsible for high-frequency real-time flight status tracking. Its core task is to convert the target waypoints and speed instructions provided by the trajectory planning layer into low-level control instructions that the drone can directly execute. The workflow begins with error calculation. The dynamic control layer continuously compares the desired state of the target position and velocity from the trajectory planning layer with the actual state from the onboard sensors, thereby generating a real-time position and velocity error vector. Subsequently, the internal control law of the dynamic control layer calculates the three-dimensional vector of the resultant force and torque that must be applied to correct the error and track the desired trajectory based on this error vector; Finally, the system decomposes this abstract force and torque vector into specific, standardized control instructions; Execution layer: The bottom execution layer receives attitude and thrust commands from the dynamic control layer and directly converts them into pulse width modulation (PWM) signals that control the speed of each motor, thereby driving the drone to complete actual physical actions; Step S3-2: Design an adaptive control gain adjustment mechanism to dynamically adjust control parameters according to flight status and environmental interference level; Step S3-3: Use a multi-level security mechanism to perform global path planning.

7. The method for controlling a UAV flight with wireless interference prevention according to claim 6, wherein: In step S3-3, a multi-level security mechanism is used to perform global path planning, specifically: In the trajectory planning layer, three progressive safety planning modes are set to deal with different levels of radio interference and navigation signal loss risks; Level 1 Normal Mode: When there is no interference or low-level interference and the main navigation signal is normal, the path planner executes the mission according to the preset route or the optimal economic route; Level 2 Avoidance Mode: This mode is activated in the event of medium to high interference intensity. Based on the possible location of the interference source or the affected area, the interference source is marked as a high-cost "virtual obstacle" in the path planning algorithm's cost map and a safe, cost-effective route is automatically replanned to bypass the area. Level 3 self-preservation mode: This mode is activated when there is extreme interference and the primary navigation signal is completely lost. The system immediately terminates the current mission and switches to the backup autonomous navigation mode to execute the preset emergency plan.

8. The method for controlling a UAV flight with wireless interference prevention according to claim 1, wherein: In step S4, a multimodal communication link and a data temporary storage mechanism are used to ensure real-time monitoring of the task status in extreme interference scenarios, specifically: Step S4-1: compress and encode the original sensor data and store them in a hierarchical manner according to their importance, specifically: Level 1 is core flight safety data including attitude, position, speed and control mode; Level 2 is critical data generated by mission payloads; Level 3 is routine equipment telemetry data; A unified, efficient, lossless compression algorithm is used to compress all classified data. The compressed data is then temporarily stored in an onboard circular buffer. The buffer adopts a hierarchical storage strategy, allocating protected, last-overwritten storage areas for primary data. This ensures that even when communications are interrupted and storage space is limited, the most critical flight records of the drone are preserved with the highest priority and in the most complete manner. Step S4-2: Adopt an adaptive channel selection mechanism to continuously evaluate the quality of each communication channel and prioritize the most reliable channel for data transmission. Specifically: The adaptive channel selection mechanism starts a timer that is set to trigger every 100 milliseconds to drive the loop execution of the entire selection process; At the beginning of each cycle, all available communication links are queried in parallel, and the original values ​​of the three key performance indicators (KPIs) of each link at the current moment, namely, the received signal strength indicator RSSI, the signal-to-noise ratio SNR, and the packet loss rate PLR, are obtained; For each link i, the collected raw RSSI value is converted into i and SNR i Convert to a standardized value in the [0,1] interval; The comprehensive quality score Q of each link is calculated through the preset weighted formula i : Q i =w R ·norm(RSSI i )+w S ·norm(SNR i )-w P ·PLR i ; Among all the link quality scores Q, the link with the highest score is determined. When its score is greater than the current link score, the link switching procedure is immediately executed to switch the main communication task to the link with the highest score. Step S4-3: Execute a state smooth transition algorithm to gradually fuse the state estimation in the emergency mode with the recovered satellite positioning data.

9. The method for controlling a UAV flight with wireless interference prevention according to claim 8, characterized in that: In step S4-3, a state smooth transition algorithm is executed to gradually fuse the state estimation in the emergency mode with the recovered satellite positioning data. The state smooth transition algorithm adopts a linear weighted fusion algorithm based on a time window to achieve smooth transition. The specific implementation is as follows: After confirming that the GNSS signal has stabilized, the process is triggered and a transition time window with a fixed duration of T is started. The start time t0 is recorded. In each control cycle within the time window, the following operations are performed: First, read the latest autonomous navigation position P VIO (t), and the current GNSS position P GNSS (t); According to the current time t, calculate a weight factor that increases linearly from 0 to 1 By weighting, the output position command P(t) that should be used in the current cycle is calculated: P(t)=(1-β(t))·P VIO (t)+β(t)·P GNSS (t); The output position command P(t) calculated in the previous step is used as the final position target of the current control cycle and sent to the underlying flight controller for execution. When time t reaches or exceeds t0+T, the transition time window ends. The satellite position P GNSS (t) is directly used as the output position instruction, and this mode is continuously used until the next navigation source is switched; the entire smooth transition process is now complete.

Images