Unmanned aerial vehicle dynamic disturbance awareness and adaptive collision avoidance method

Abstract

The present application belongs to the technical field of low-altitude geological detection and unmanned aerial vehicle flight control, and specifically relates to a method for dynamic disturbance perception and self-adaptive anti-collision of an unmanned aerial vehicle, which collects in real time the attitude, flight state, terrain parameters and measured wind field data of the unmanned aerial vehicle; inputs the tensor into a double-encoder wind field prediction neural network, a spatial encoder extracts spatial features of the terrain, a time-series encoder extracts time-series evolution features, and the two features are fused through cross-attention to output predicted disturbance intensity and predicted confidence; when the confidence is lower than a threshold, a self-supervised training signal is generated by repeated observation of the same geographical location on different flight paths; a comprehensive risk coefficient is constructed according to the predicted disturbance intensity and the confidence, and an adaptive avoidance route based on an improved A-star algorithm is triggered when the risk exceeds a limit. The present application realizes online self-adaptation of the model without external labeling, and significantly improves the flight safety of low-altitude geological detection unmanned aerial vehicles in complex terrain environments.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) geological exploration and flight control technology, and in particular to a method for UAV dynamic disturbance perception and adaptive collision avoidance. Background Technology

[0002] Current drone collision avoidance technologies primarily rely on visual or lidar detection of physical obstacles, falling under the category of "reactive" obstacle avoidance. These methods cannot anticipate the invisible airflow threats induced by terrain. By the time a drone enters a leeward slump or a turbulent ridge zone, it is often too late to correct its course. For example, Chinese patent application CN115285338A only considers the simple impact of wind direction on takeoff and landing. Chinese patent application CN119045524B proposes a formation obstacle avoidance method based on a dynamic perturbation fluid system, but its perturbation sources mainly target static or moving physical obstacles, failing to consider terrain-induced dynamic airflow fields as predictable perturbation sources. While Chinese patent application CN119828725A involves aerodynamic interference, its interference sources are based on swarm motion simulations and are not applicable to single-drone low-altitude detection.

[0003] Therefore, there is an urgent need for advanced collision avoidance methods capable of real-time perception, quantitative prediction, and proactive avoidance of dynamic airflow threats. However, achieving this goal faces three contradictory technical challenges: First, terrain-induced disturbances are characterized by strong spatial non-uniformity and suddenness, requiring prediction models to possess high accuracy and rapid response capabilities; second, different terrain features vary greatly, making it difficult for offline-trained models to adapt to changes in terrain conditions during flight, necessitating online adaptive capabilities; third, UAVs have extremely limited onboard computing resources, making them unable to withstand frequent full-model retraining. Existing technologies either only address physical obstacles (reactive obstacle avoidance) or treat airflow as unpredictable external disturbances (passive response), never realizing that terrain-induced disturbances can be treated as a predictable spatial distribution field, and leveraging the inherent serpentine scanning trajectory repetitive observation characteristics in geological survey tasks to achieve online adaptive prediction models with extremely low computational overhead. Summary of the Invention

[0004] This invention proposes a dynamic disturbance perception and adaptive collision avoidance method for UAVs, aiming to solve the technical problems of slow response and inability to predict invisible airflow threats in existing UAV obstacle avoidance technologies under complex airflow environments.

[0005] This invention provides a method for dynamic disturbance perception and adaptive collision avoidance of unmanned aerial vehicles (UAVs), including: S101: Acquire UAV attitude data, flight status data, terrain parameters, and airborne measured wind field data; after aligning by timestamp, stitch them together along the channel dimension to construct an input tensor with a time step dimension and a feature channel number dimension; S102, the input tensor is input into a dual-encoder wind field prediction neural network, the dual-encoder wind field prediction neural network comprising: a spatial encoder for extracting spatial distribution features of terrain from the input tensor; a temporal encoder for extracting temporal evolution features of attitude data and wind field data from the input tensor; a cross-attention fusion module for performing feature interaction fusion of the outputs of the spatial encoder and the temporal encoder; and an output decoder for decoding the predicted turbulence intensity and prediction confidence of future moments from the fused features. S103, during the process of the UAV performing S-shaped scanning flight, it is determined whether the prediction confidence is lower than a preset first threshold. If so, a self-supervised training signal is generated using repeated observation data of the same geographical location on adjacent tracks. The mean square error between the predicted turbulence intensity and the measured turbulence intensity is used as the loss function to update the parameters of the batch normalization layer in the dual encoder wind field prediction neural network online, while keeping the weight parameters of the convolutional layer and the fully connected layer in the network unchanged. S104, Construct a comprehensive flight risk coefficient based on the predicted turbulence intensity and the predicted confidence level, wherein the comprehensive flight risk coefficient is positively correlated with the predicted turbulence intensity and negatively correlated with the predicted confidence level; S105, determine whether the comprehensive flight risk coefficient is greater than the preset second threshold. If so, generate an adaptive avoidance route and execute it.

[0006] Further, S101 specifically includes: The UAV attitude data is arranged in a 6-dimensional vector form of pitch angle, roll angle, yaw angle and their respective angular velocities; the flight status data is arranged in a 3-dimensional vector form of velocity, altitude and acceleration; the terrain parameters are arranged in a 2-dimensional vector form of slope angle and aspect angle; and the airborne measured wind field data is arranged in a 4-dimensional vector form of three-dimensional wind speed components and turbulence intensity. The four types of data are spliced ​​along the channel dimension in a fixed order of attitude data, flight status data, terrain parameters and wind field data at the same sampling time to form a 15-dimensional feature vector at that time. The 15-dimensional feature vectors of all sampling times within a continuous T seconds are stacked along the time dimension to form the input tensor.

[0007] Furthermore, the temporal encoder employs a multi-layer one-dimensional temporal convolutional network, with each temporal convolutional layer containing 32 or 64 convolutional kernels, a kernel size of 3 or 5, and a stride of 1; the spatial encoder employs a two-dimensional convolutional neural network, taking a two-dimensional feature map composed of terrain parameters in the spatial dimension as input, and extracting local spatial patterns and global structural features of the terrain through multiple convolutional layers. The cross-attention fusion module uses the output of the temporal encoder as the query vector and the key vector and value vector obtained by linear projection of the output of the spatial encoder as the key-value pair vector. It calculates the correlation weight between the temporal features and the spatial features through the attention mechanism, and then concatenates the spatial features with the temporal features based on the correlation weight to form the fused features.

[0008] Furthermore, the output of the prediction confidence branch in the output decoder is equipped with a temperature scaling module. The temperature scaling module scales the input of the Softmax function through a learnable temperature parameter, so that the prediction confidence can reflect the model's familiarity with the current input sample.

[0009] Furthermore, in step S103, generating a self-supervised training signal using repeated observation data from adjacent tracks at the same geographical location specifically includes: During S-shaped scan flight, the same geographical location is successively covered by sensors of adjacent flight paths. The sensor data when the location was passed for the first time is used as the input sample, and the measured wind field data when the location was passed for the second time is used as the supervision label to form a self-supervised training sample pair. The loss function of the self-supervised training signal adopts the mean square error between the predicted turbulence intensity and the measured turbulence intensity.

[0010] Furthermore, in S103, the parameters of the batch normalization layer in the dual-encoder wind field prediction neural network are updated online, specifically including: The scaling factor and offset factor of the batch normalization layer in the dual encoder wind field prediction neural network are updated, while the weight parameters of the convolutional layer and the fully connected layer remain unchanged. When the number of accumulated self-supervised training samples reaches the preset batch size, a gradient descent update is performed, and the sample buffer is cleared after the update is completed.

[0011] Furthermore, the method for constructing the comprehensive flight risk coefficient in S104 is as follows: the predicted turbulence intensity is normalized to obtain the normalized turbulence intensity, the prediction confidence is transformed by negative logarithm to obtain the uncertainty penalty term, and the normalized turbulence intensity and the uncertainty penalty term are weighted and summed and then added to the preset benchmark risk coefficient to obtain the comprehensive flight risk coefficient.

[0012] Furthermore, the method for generating an adaptive avoidance path in S105 is as follows: starting from the current position and ending at a preset safe return point or mission target point, a path search is performed in three-dimensional space using the A* algorithm; during the search process of the A* algorithm, each candidate node is verified for feasibility based on the UAV's kinematic model. This verification includes, but is not limited to, determining whether the curvature of the path formed by consecutive nodes exceeds the UAV's minimum turning radius and whether the path climb rate exceeds the maximum climb rate.

[0013] Furthermore, the low-altitude geological detection task includes at least one of the following: airborne magnetic survey of mineral resources, airborne gamma-ray spectroscopy measurement, hyperspectral mineral mapping, and inspection of potential geological disaster sites; the detection sensors carried by the UAV include at least one of the following: airborne magnetometer, airborne gamma-ray spectrometer, hyperspectral imager, and synthetic aperture radar.

[0014] Furthermore, the method also includes S106: packaging and uploading the sensor data collected in each flight mission, the online update records of the dual encoder wind field prediction neural network, and the flight logs to the cloud server to build a cross-mission terrain-disturbance database, providing a reference for model initialization of subsequent flight missions.

[0015] Technical effects: (1) By using a dual encoder structure, the terrain spatial features and flight state temporal features are extracted and then fused separately, which significantly improves the prediction accuracy compared to a single network; (2) By using a confidence branch with temperature scaling, a quantifiable reliability index is provided for turbulence prediction for the first time, so that subsequent online learning is only triggered when the model is "uncertain", effectively avoiding the degradation of model performance (catastrophic forgetting) caused by blind updates; (3) By using an efficient strategy of updating only the parameters of the batch normalization layer, the computational cost of online adaptation is reduced by two orders of magnitude (compared to full model update), thus achieving real-time feasibility on an airborne embedded platform; (4) By using the predicted turbulence intensity as a weight factor of the path cost of the A-star algorithm, a leap from "passive obstacle avoidance" to "active turbulence avoidance" is achieved, and in conjunction with confidence assessment, even when the predicted turbulence is moderate but the model is uncertain, avoidance will be triggered, significantly improving flight safety. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the UAV dynamic disturbance perception and adaptive collision avoidance method proposed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the dual-encoder wind field prediction neural network in this invention; Figure 3 This is a schematic diagram of the temporal relationship of the S-shaped scanning self-supervised training in this invention; Figure 4 This is a flowchart illustrating the adaptive avoidance route generation method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This paper addresses the technical problems mentioned in the background technology regarding the inability of low-altitude geological inspection UAVs to adapt their prediction models online under airflow disturbances induced by complex terrain, the lack of prediction confidence assessment, and the inability to utilize the inherent repeated observation characteristics of serpentine scanning tracks for self-supervised learning.

[0019] It should be noted that the "repeated observation" upon which this invention relies naturally exists within the standard serpentine scanning flight path of geological detection tasks. This is a key premise that distinguishes this invention from other UAV application scenarios. Serpentine scanning not only covers the detection area but also provides a valuable self-supervised signal source for the prediction model. This invention cleverly combines task flight path design with continuous model learning, achieving a closed loop of "flight as training".

[0020] To achieve the above objectives, this invention provides a method for dynamic disturbance perception and adaptive collision avoidance using a UAV. This embodiment uses a UAV equipped with a hyperspectral imager and LiDAR performing a mineral resource exploration mission in a mountainous area as an example. This area has complex terrain features, including windward slopes, leeward slopes, and ridgelines. Snake scanning refers to the UAV flying back and forth over the detection area according to a preset "S"-shaped flight path, maintaining a fixed horizontal distance between adjacent flight paths to achieve comprehensive coverage of the target area. (Reference) Figure 1 As shown, the specific steps include: Step S101: Multidimensional time series data acquisition and input tensor construction.

[0021] During low-altitude geological surveys using a serpentine scanning flight, the UAV collects four types of multi-dimensional time-series data in real time at a frequency of 20Hz via onboard sensors. These four types of data are then aligned by timestamps and constructed into an input tensor that can be processed by a neural network. The physical relationships between these four types of data—namely, terrain determining airflow patterns through spatial distribution, attitude reflecting airflow effects through temporal responses, and flight status characterizing the degree of disturbance to the UAV—will be captured and fused separately using a subsequent dual-encoder neural network structure.

[0022] (1) UAV attitude data: including pitch angle, roll angle, yaw angle and corresponding angular velocity, totaling 6 dimensions. The data comes from the onboard inertial measurement unit (IMU), with a sampling frequency of 100Hz. The raw data is obtained after processing by extended Kalman filtering. This data reflects the attitude response of the UAV under the current airflow disturbance and is a direct external characterization of the disturbance intensity.

[0023] (2) Flight status data: including flight speed, altitude, and acceleration, totaling 3 dimensions. The data comes from the fusion results of the GPS / INS integrated navigation system and the barometric altimeter. This data reflects the macroscopic motion state of the UAV and together with the attitude data constitutes the complete motion response under the influence of turbulence.

[0024] (3) Terrain parameters: including the slope angle and aspect angle of the terrain below the UAV, totaling 2 dimensions. The data comes from point cloud data generated by real-time scanning by airborne LiDAR. The slope angle (the angle between the plane normal vector and the horizontal plane) and aspect angle (the angle between the direction of the fastest descent of the terrain and due north) are calculated within each terrain profile window. This data is the prior spatial information for turbulence prediction—terrain is the root cause of turbulence, and the slope angle and aspect angle determine the spatial distribution of airflow lifting or sinking.

[0025] (4) Airborne measured wind field data: including three-dimensional wind speed components (u component, v component, w component) and turbulence intensity, totaling 4 dimensions. The data comes from the multi-source fusion results of the nose micro five-hole probe and the wing leading edge MEMS wall shear stress sensor. This data is a direct measurement of the current turbulence, serving as both one of the input features of the neural network and a source of supervision labels in self-supervised training.

[0026] The four types of data are aligned on the timestamp and arranged in a fixed channel order. Specifically, the data vector at each sampling time t is: 6-dimensional attitude, 3-dimensional flight state, 2-dimensional terrain parameters, and 4-dimensional wind field data, for a total of 15 dimensions. Data from 5 consecutive seconds (sampling frequency 20Hz, 100 time steps) are stacked along the time dimension to construct an input tensor of shape 100×15. This tensor serves as the input to the subsequent dual-encoder wind field prediction neural network.

[0027] ; in, This is 6-dimensional attitude data. For 3D flight status data, For 2D terrain parameters, This is 4D wind field data.

[0028] Data from consecutive T seconds (T=5 seconds, sampling frequency 20Hz, 100 time steps) are collected and stacked along the time dimension to construct an input tensor of shape (100,15). This tensor serves as the input to the subsequent dual-encoder wind field prediction neural network.

[0029] Step S102, inference of the dual-encoder wind field prediction neural network. (Refer to...) Figure 2This step employs a pre-trained dual-encoder wind field prediction neural network to perform real-time inference on the input tensor constructed in step S101, outputting the predicted turbulence intensity and prediction confidence for future timeframes. The design principle of the dual-encoder structure lies in the fundamental differences in physical properties between terrain parameters and wind field / attitude data—terrain parameters are spatially distributed data, while attitude and wind field data are temporally sequenced data; the two types of data are not suitable for processing with the same encoder. Therefore, this invention designs a spatial encoder and a temporal encoder to process the data separately, and then achieves deep fusion of the two types of features through a cross-attention module.

[0030] Specifically, the network is pre-trained end-to-end using time-series data (terrain parameters, attitude, and wind field) collected from historical flight missions, with measured wind field data for a future period as supervision labels, and employing standard supervised learning methods (such as the Adam optimizer and MSE loss function). It can even be mentioned that pre-training can be performed on general terrain data with similar geographical features.

[0031] The spatial encoder is used to extract the spatial distribution features of the terrain. Specifically, the terrain parameter dimensions (slope angle and aspect angle) in the input tensor are extracted from the time series and reconstructed into a two-dimensional spatial feature map with a size of 10×10 grid. The spatial encoder uses a two-dimensional convolutional neural network: the first convolutional layer uses a 5×5 convolutional kernel and outputs 32 channels to extract the local spatial patterns of the terrain; the second convolutional layer uses a 3×3 convolutional kernel and outputs 64 channels to extract the global structural features of the terrain; finally, a flattening layer outputs a 6400-dimensional feature vector.

[0032] The temporal encoder is used to extract temporal evolution features of attitude and wind field data from the input tensor. The temporal encoder employs a multi-layer one-dimensional temporal convolutional network: the first layer uses a size 5 convolutional kernel and outputs 32 channels to capture attitude and wind field changes over short timescales; the second layer uses a size 3 convolutional kernel and outputs 64 channels to capture evolution over longer timescales; finally, global average pooling is used to output a 64-dimensional feature vector.

[0033] Regarding the cross-attention fusion module and the calculation of the attention weight matrix: The temporal encoder outputs a 64-dimensional feature vector, denoted as the query vector Q, which represents the current temporal evolution state of the UAV attitude and wind field. The spatial encoder outputs a 6400-dimensional feature vector, which is first mapped to a key vector K and a value vector V through two independent linear projection matrices, reducing both dimensions to 64. The key vector K is used to calculate the relevance weights with the query vector Q, and the value vector V carries the specific information of the spatial features.

[0034] The formula for calculating the attention weight matrix A is: ; Where Q is the query vector output by the temporal encoder, with a shape of 1×64, representing the temporal features at the current time step, and K is the key vector output by the spatial encoder after linear projection, with a shape of 1×64, representing the index of the terrain spatial features. This represents the dot product of the query vector and the transpose of the key vector, resulting in a scalar that measures the correlation between the current temporal features and the terrain spatial features. The larger the dot product value, the stronger the correlation between the two. This is the dimension of the key vector, which is 64 in this case. (Use...) Scaling is performed to prevent the dot product value from becoming too large, causing the softmax function to enter the gradient saturation region and affecting the training effect. A is the calculated attention weight matrix, with a 1×1 shape (scalar weights), representing the degree of attention the current temporal feature pays to the overall terrain spatial features. Subsequently, the attention weights A are multiplied by the value vector V to obtain the weighted spatial feature: Z = AV. This weighted feature represents the part of the terrain that contributes most to turbulence under the current flight attitude and wind field conditions.

[0035] Finally, the weighted spatial feature Z (64-dimensional) is concatenated with the original temporal feature (64-dimensional) output by the temporal encoder along the channel dimension to form a 128-dimensional fused feature vector. This fused vector contains both temporal and spatial information for use by the subsequent decoder.

[0036] The output decoder consists of two fully connected layers. The input is a 128-dimensional fused feature vector, and the output is divided into two branches: the first branch outputs the predicted turbulence intensity sequence for 40 time steps within the next 2 seconds, expressed in cubic seconds per square meter; the second branch outputs the prediction confidence score, ranging from 0 to 1. A temperature scaling module is incorporated at the output of the confidence score branch. This module uses a learnable temperature parameter to scale the input to the Softmax function, ensuring that the prediction confidence score reflects the model's familiarity with the current input sample.

[0037] During the flight of the UAV, the onboard mission computer performs inference at a frequency of 20Hz: it reads the sensor data of the most recent 5 seconds from the data buffer to construct the input tensor, sends it to the neural network for forward propagation, and obtains the predicted turbulence intensity and prediction confidence.

[0038] Step S103: Online self-supervised training based on confidence assessment.

[0039] This step is one of the core innovations that distinguishes this invention from existing technologies. Existing deep learning methods use offline training modes with fixed model parameters, which cannot adapt to changes in terrain features during flight. The core insight of this invention lies in the fact that in the serpentine scanning tracks of geological detection, the same geographical location is successively covered by sensors on adjacent tracks, which naturally provides self-supervised training signals without external annotation.

[0040] Reference Figure 3 The principle of self-supervised training for serpentine scanning (i.e., S-shaped scanning) is as follows: (1) During the serpentine scanning flight, the UAV covers the detection area according to the preset "S" shaped route. The horizontal distance between adjacent tracks is D=35 meters, and the flight speed is V=8 meters / second. Therefore, the time difference Δt≈D / V≈4.4 seconds between adjacent track sensors covering the same geographical location is approximately 4.4 seconds.

[0041] (2) When the UAV flies along the flight path A and passes through the geographical location P, it collects sensor data as input samples. The predicted turbulence intensity is obtained through neural network inference. .

[0042] (3) When the UAV flies over the same geographical location P again along the flight path B (with a time difference of about 4.4 seconds), the measured wind field data at that location is collected as a supervision label. Since the time interval between the two passages is very short (4.4 seconds is much smaller than the timescale of weather system evolution, which typically lasts for several hours or more), it can be assumed that the wind field conditions remain basically unchanged during this period. Can be used as The corresponding true value of the turbulence.

[0043] (4) Combine sample-label pairs [ , The samples are stored in the self-supervised sample buffer. When the number of samples in the buffer reaches the preset batch size (32 samples in this embodiment), a gradient descent update is performed.

[0044] To further verify the rationality of the assumption that "the wind field remains unchanged over a short period of time," it is necessary to analyze the physical characteristics and time scale of terrain-induced turbulence. According to fluid mechanics theory, the associated length (integral scale) of turbulent vortices induced by changes in terrain roughness and slope is typically on the order of the size of the terrain feature (e.g., a ridge height of 100m). The time scale of turbulent energy cascade τ≈L / U, where L is the integral scale and U is the average wind speed (approximately 10m / s), thus τ≈10 seconds. The time difference between adjacent tracks Δt≈4.4 seconds is smaller than this characteristic time scale; therefore, under engineering approximation, the turbulence field can be considered quasi-steady-state. Simultaneously, this invention designs an environmental consistency verification mechanism: when generating self-supervised sample pairs, the average wind speed and wind direction measured by the airborne anemometer are compared between the two observations. If the wind speed change exceeds 15% or the wind direction change exceeds 20 degrees, it is determined that the environmental conditions have changed significantly, and the sample pair is discarded and not used for training. This mechanism further ensures the reliability of the self-supervised labels.

[0045] The loss function for self-supervised training uses the mean square error between the predicted and measured turbulence intensity: ; Where N is the batch size.

[0046] Online parameter updates employ an efficient update strategy: only the scaling factor γ and offset factor β of the batch normalization layer in the dual-encoder wind field prediction neural network are updated, while the weight parameters of the convolutional and fully connected layers remain unchanged. The rationale for this strategy is that the convolutional and fully connected layers learn the fundamental mapping relationship between terrain and airflow, exhibiting good generalization ability and requiring less frequent updates; while the parameters of the batch normalization layer reflect the statistical distribution characteristics of the current input data, enabling rapid adaptation to changes in terrain features, and have a very small number of parameters (only 2 parameters per layer), resulting in extremely low computational overhead for updates. The update formula for the batch normalization layer is standard gradient descent: ; ; in The learning rate is set to 0.001 in this example. After the update is complete, the self-supervised sample buffer is cleared, and the next accumulation continues.

[0047] Compared to large-scale weather systems (several hours), the characteristic evolution timescale of topographically induced dynamic disturbances (such as leeward vortices and mountain wave turbulence) is typically on the order of minutes or even longer. The selected time difference between adjacent tracks (e.g., 4.4 seconds) is much smaller than this characteristic timescale, and therefore can be treated as quasi-static.

[0048] A sample is only used for self-supervised training if the change in environmental parameters (such as average wind speed and wind direction) between two observations is less than a certain proportion; otherwise, it is discarded. This enhances the robustness and reliability of the technical solution.

[0049] Step S104: Construction of Comprehensive Flight Risk Coefficient. This step uses the predicted turbulence intensity output from step S102. and prediction confidence It is combined into a single risk coefficient to determine whether to trigger a collision avoidance action.

[0050] First, predict the turbulence intensity Normalize to the [0,1] interval. Let the known maximum turbulence intensity be... =0.5m² / s³, then the normalized turbulence intensity is =min( / ,1).

[0051] Secondly, the prediction confidence level This is transformed into an uncertainty penalty term. The lower the confidence level (the more uncertain the model), the larger the penalty term. A negative logarithmic transformation is used. =-ln( +δ), where δ=1e-6 is the amount to prevent the removal of zero. When When =0.5, ≈0.69; when When =0.1, ≈2.30.

[0052] Finally, the weighted sum of the normalized turbulence intensity and the uncertainty penalty term is compared with the preset benchmark risk coefficient. Adding 0.3 together, we get the comprehensive flight risk coefficient: ; in =0.4, =0.3. The principle behind this design is that the determination of a high-risk zone depends not only on the intensity of the disturbance itself, but also on the certainty of the model's prediction for that area—even if the predicted disturbance intensity is moderate, if the model confidence is extremely low (indicating that the terrain-airflow combination is not learned by the model), a collision avoidance action should still be triggered.

[0053] Step S105: Adaptive evasion path generation. (Refer to...) Figure 4 When the comprehensive flight risk coefficient calculated in step S104 Greater than the preset second threshold When the value is 0.8, it is determined to be a high-risk state, and the collision avoidance action is triggered.

[0054] This step uses an improved A-star (A The algorithm performs path search in a three-dimensional discrete space. The specific implementation is as follows: (1) State Space and Node Definition: Establish a three-dimensional rectangular coordinate system with the UAV takeoff point or mission start point as the origin, where the X-axis points due east, the Y-axis points due north, and the Z-axis vertically upwards represents altitude. Divide the planned area in front of the UAV (e.g., 200m horizontally in front and behind, 200m horizontally to the left and right, and -30m to +50m vertically relative to the current altitude) into a uniform grid with a grid resolution of 5m. Let the three-dimensional coordinates of any candidate grid node n be... The three-dimensional coordinates of the preset endpoint (safe return point or mission objective point) G are: The current starting point is .

[0055] (2) Path cost function: In the A* algorithm, the total cost f(n) of each node n consists of two parts: the actual cost g(n) and the heuristic estimate h(n), i.e. f(n) = g(n) + h(n).

[0056] Here, the actual cost g(n) is defined as the cumulative path cost from the starting point to node n. Let Δs be the parent node p (coordinates are...). The three-dimensional straight-line distance between node n and the current candidate node n is calculated using the following formula: .

[0057] The single-step cost from parent node p to current node n is: ,in The predicted disturbance intensity at the location of node n is obtained by mapping the node's coordinates to the terrain grid and querying the neural network output. =2.0 is the turbulence cost weighting coefficient. Therefore, the recursive formula for the actual cost is: .

[0058] This design causes the path to tend to traverse areas of low turbulence intensity, even with a slight increase in path length.

[0059] The heuristic estimate h(n) is used to estimate the minimum possible cost from the current node n to the target point G. It employs three-dimensional Euclidean distance and superimposes a turbulence penalty term to guide the search direction while balancing distance and safety. ; in =1.5 is the heuristic perturbation weight.

[0060] (3) Kinematic Feasibility Verification: For each three-point path segment consisting of parent node p, current node n, and next candidate node q, it is necessary to verify whether it satisfies the kinematic constraints of the UAV. The verification rules are as follows: Minimum turning radius verification: Determine the radius R of the circumcircle using three points p, n, and q. Let the lengths of the three sides of the triangle be a = |nq|, b = |pq|, and c = |pn|, and the area of ​​the triangle be... (Calculated using Heron's formula), then the radius of the circumcircle... .like (In this embodiment) If the curvature of the path exceeds the minimum turning radius of the UAV (=15m), then the node q is discarded.

[0061] Maximum climb / glide rate verification: Calculate the height difference between the current node n and the next candidate node q. And horizontal distance: ; like (in If the maximum pitch angle of the drone is 20 degrees in this embodiment, then the climb or descent rate is determined to be out of limit, and node q is discarded.

[0062] (4) Search Termination and Backup Strategy: When the extended node reaches the neighborhood of the target point G (i.e., the three-dimensional Euclidean distance is less than 5m of the grid resolution), the parent pointer is backtracked to generate a complete avoidance route and output it. If no feasible path is found within the preset maximum number of iterations (5000 times in this embodiment), the backup safety strategy is triggered: the UAV performs a vertical climb to the preset safe altitude (50m above the starting point) and then flies straight to the target point.

[0063] Step S106: Cloud Data Aggregation and Cross-Mission Knowledge Transfer. Sensor data collected during each flight mission, online update records of the dual-encoder wind field prediction neural network (including the location of the triggered update and the updated batch normalization layer parameters), and flight logs are packaged and uploaded to the cloud server. The cloud server aggregates data from multiple UAVs under different terrain conditions to build a cross-mission terrain-disturbance database. When a UAV performs a new geological survey mission, it can retrieve historical flight data from the cloud server that best matches the terrain features of the current mission area. This data is used to initialize the dual-encoder wind field prediction neural network, thereby reducing the time required for online adaptation.

[0064] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method for dynamic disturbance perception and adaptive collision avoidance of unmanned aerial vehicles (UAVs), characterized in that, include: S101: Acquire UAV attitude data, flight status data, terrain parameters and airborne measured wind field data, align them by timestamp, and then stitch them along the channel dimension to construct an input tensor with time step dimension and feature channel number dimension. S102, the input tensor is input into a dual-encoder wind field prediction neural network, the dual-encoder wind field prediction neural network comprising: a spatial encoder for extracting spatial distribution features of terrain from the input tensor; a temporal encoder for extracting temporal evolution features of attitude data and wind field data from the input tensor; a cross-attention fusion module for performing feature interaction fusion of the outputs of the spatial encoder and the temporal encoder; and an output decoder for decoding the predicted turbulence intensity and prediction confidence of future moments from the fused features. S103, during the process of the UAV performing S-shaped scanning flight, it is determined whether the prediction confidence is lower than a preset first threshold. If so, a self-supervised training signal is generated using repeated observation data of the same geographical location on adjacent tracks. The mean square error between the predicted turbulence intensity and the measured turbulence intensity is used as the loss function to update the parameters of the batch normalization layer in the dual encoder wind field prediction neural network online, while keeping the weight parameters of the convolutional layer and the fully connected layer in the network unchanged. S104, Construct a comprehensive flight risk coefficient based on the predicted turbulence intensity and the predicted confidence level, wherein the comprehensive flight risk coefficient is positively correlated with the predicted turbulence intensity and negatively correlated with the predicted confidence level; S105, determine whether the comprehensive flight risk coefficient is greater than the preset second threshold. If so, generate an adaptive avoidance route and execute it.

2. The method according to claim 1, characterized in that, S101 specifically includes: The UAV attitude data is arranged in a 6-dimensional vector form of pitch angle, roll angle, yaw angle and their respective angular velocities; the flight status data is arranged in a 3-dimensional vector form of velocity, altitude and acceleration; the terrain parameters are arranged in a 2-dimensional vector form of slope angle and aspect angle; and the airborne measured wind field data is arranged in a 4-dimensional vector form of three-dimensional wind speed components and turbulence intensity. The four types of data are spliced ​​along the channel dimension in a fixed order of attitude data, flight status data, terrain parameters and wind field data at the same sampling time to form a 15-dimensional feature vector at that time. The 15-dimensional feature vectors of all sampling times within a continuous T seconds are stacked along the time dimension to form the input tensor.

3. The method according to claim 1, characterized in that, The temporal encoder employs a multi-layer one-dimensional temporal convolutional network, with each temporal convolutional layer containing 32 or 64 convolutional kernels, a kernel size of 3 or 5, and a stride of 1. The spatial encoder employs a two-dimensional convolutional neural network, taking a two-dimensional feature map composed of terrain parameters in the spatial dimension as input, and extracting local spatial patterns and global structural features of the terrain through multiple convolutional layers. The cross-attention fusion module uses the output of the temporal encoder as the query vector and the key vector and value vector obtained by linear projection of the output of the spatial encoder as the key-value pair vector. It calculates the correlation weight between the temporal features and the spatial features through the attention mechanism, and then concatenates the spatial features with the temporal features based on the correlation weight to form the fused features.

4. The method according to claim 1, characterized in that, The output of the prediction confidence branch in the output decoder is equipped with a temperature scaling module. The temperature scaling module scales the input of the Softmax function through a learnable temperature parameter, so that the prediction confidence can reflect the model's familiarity with the current input sample.

5. The method according to claim 1, characterized in that, S103, which generates a self-supervised training signal using repeated observation data from adjacent tracks at the same geographical location, specifically includes: During S-shaped scan flight, the same geographical location is successively covered by sensors of adjacent flight paths. The sensor data when the location was passed for the first time is used as the input sample, and the measured wind field data when the location was passed for the second time is used as the supervision label to form a self-supervised training sample pair. The loss function of the self-supervised training signal adopts the mean square error between the predicted turbulence intensity and the measured turbulence intensity.

6. The method according to claim 1, characterized in that, In step S103, the parameters of the batch normalization layer in the dual-encoder wind field prediction neural network are updated online, specifically including: The scaling factor and offset factor of the batch normalization layer in the dual encoder wind field prediction neural network are updated, while the weight parameters of the convolutional layer and the fully connected layer remain unchanged. When the number of accumulated self-supervised training samples reaches the preset batch size, a gradient descent update is performed, and the sample buffer is cleared after the update is completed.

7. The method according to claim 1, characterized in that, The method for constructing the comprehensive flight risk coefficient in S104 is as follows: the predicted turbulence intensity is normalized to obtain the normalized turbulence intensity, the prediction confidence is transformed by negative logarithm to obtain the uncertainty penalty term, and the normalized turbulence intensity and the uncertainty penalty term are weighted and summed and then added to the preset benchmark risk coefficient to obtain the comprehensive flight risk coefficient.

8. The method according to claim 1, characterized in that, The method for generating an adaptive avoidance path in S105 is as follows: starting from the current position and ending at a preset safe return point or mission target point, the A* algorithm is used to search for a path in three-dimensional space; during the search process of the A* algorithm, each candidate node is verified for feasibility based on the UAV's kinematic model. This verification includes, but is not limited to, determining whether the curvature of the path formed by consecutive nodes exceeds the UAV's minimum turning radius and whether the path climb rate exceeds the maximum climb rate.

9. The method according to claim 1, characterized in that, The low-altitude geological detection mission includes at least one of the following: airborne magnetic survey of mineral resources, airborne gamma-ray spectroscopy measurement, hyperspectral mineral mapping, and inspection of potential geological disaster sites; the detection sensors carried by the UAV include at least one of the following: airborne magnetometer, airborne gamma-ray spectrometer, hyperspectral imager, and synthetic aperture radar.

10. The method according to claim 1, characterized in that, The method further includes S106: packaging and uploading the sensor data collected in each flight mission, the online update records of the dual encoder wind field prediction neural network, and the flight logs to the cloud server to build a cross-mission terrain-disturbance database, providing a reference for model initialization of subsequent flight missions.

Images