CN-122021324-A - Unmanned aerial vehicle inspection operation obstacle avoidance training simulation method

CN122021324ACN 122021324 ACN122021324 ACN 122021324ACN-122021324-A

Abstract

The invention provides an obstacle avoidance training simulation method for unmanned aerial vehicle inspection operation, which comprises the steps of obtaining a standardized obstacle state input vector through multi-step normalization and feature extraction by utilizing historical observation data, constructing a space-time memory unit through a differential nerve image machine, fusing space and time features, realizing efficient coding and storage of obstacle tracks, adopting an attention mechanism to extract local evolution trend, completing obstacle group association analysis and risk assessment by assisting a graph rolling network, predicting future space-time distribution, combining real-time observation and probability distribution, generating and screening diversified low-risk paths, outputting a lightweight action strategy, adapting to environment dynamic changes through layered knowledge distillation and an online fine tuning mechanism, improving the generalization capability and environment adaptability of an unmanned aerial vehicle obstacle avoidance strategy model, realizing self-adaptive closed-loop control, and remarkably improving unmanned aerial vehicle inspection safety and path decision efficiency.

Inventors

Yin Kengxiong
HU JING

Assignees

广州骏耀信息科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260204

Claims (10)

1. The obstacle avoidance training simulation method for the unmanned aerial vehicle inspection operation is characterized by comprising the following steps of: S1, acquiring a time sequence sample based on historical observation data of a dynamic obstacle in unmanned aerial vehicle inspection operation, and carrying out normalization processing on the time sequence sample to generate a standardized obstacle state input vector; s2, constructing a space-time memory unit by utilizing a differentiable neural turing machine structure, wherein based on the standardized obstacle state input vector, the position vector of a space channel and the relative displacement difference value label of a time channel are jointly encoded through a two-channel writing mechanism to form a memory feature matrix; s3, reading obstacle state track segments from the memory feature matrix, calculating the correlation weight of each segment and the current environment evolution mode, and generating a local evolution trend vector; s4, inputting the local evolution trend vector into a graph convolutional network, executing neighborhood aggregation operation based on topological connection relations among the obstacles, identifying potential collision clusters and main flow movement directions, and outputting probability distribution envelope surfaces of the obstacles; S5, constructing a local obstacle avoidance decision space based on the probability distribution envelope surface and the current position information of the unmanned aerial vehicle, and generating a candidate flight path set in the local obstacle avoidance decision space, wherein each path corresponds to an action strategy sequence under a group of space-time constraints; S6, taking action strategy distribution generated by the teacher network in the high-fidelity simulation environment as a soft target, carrying out layered knowledge distillation training on the boundary end student network, and outputting an obstacle avoidance strategy model.
2. The simulation method for obstacle avoidance training of unmanned aerial vehicle inspection operation according to claim 1, wherein the following step S6 further comprises: S7, collecting newly observed obstacle evolution states in real time in an actual flight process, calculating evolution residual errors between the newly observed obstacle evolution states and a predicted result, judging whether the evolution residual errors exceed a preset threshold value, and triggering an online fine tuning flow if the evolution residual errors exceed the preset threshold value; and S8, executing incremental parameter updating according to the trigger signal, and locally optimizing the obstacle avoidance strategy model by utilizing the latest observation data to generate a closed-loop obstacle avoidance control strategy with environmental adaptability.
3. The simulation method for obstacle avoidance training of unmanned aerial vehicle inspection operation according to claim 1, wherein the step S1 specifically comprises: Acquiring a dynamic obstacle historical observation data stream recorded by an onboard sensor or a simulation system in the unmanned aerial vehicle inspection process, and taking the data stream as an original observation input condition; Based on the original observation input conditions, executing data alignment and time sequence synchronization processing, filling a data missing frame caused by communication delay or sampling asynchronism by using an interpolation algorithm, resampling state sequences of all obstacles to a unified time reference, and generating time aligned obstacle state time sequence fragments; Extracting key dynamic characteristic parameters from the time-aligned obstacle state time sequence fragments to form a high-dimensional obstacle state characteristic vector; according to the high-dimensional obstacle state feature vector, numerical compression processing based on maximum-minimum normalization is carried out, and each dimensional feature is mapped into a unified interval to obtain a normalized obstacle state feature matrix; And carrying out outlier detection and robust correction on the normalized obstacle state feature matrix, removing data points which are obviously deviated from a normal motion mode by adopting an outlier identification method based on a Markov distance, replacing the data points by a weighted average value of adjacent moments, and outputting a structured normalized obstacle state input vector.
4. A simulation method for obstacle avoidance training for unmanned aerial vehicle inspection according to claim 3, wherein the dynamic obstacle history observation data stream comprises three-dimensional coordinates, linear velocity vectors, angular velocity components, acceleration modulus values, and directional derivatives thereof for each obstacle in successive time steps.
5. The simulation method for obstacle avoidance training of unmanned aerial vehicle inspection operation according to claim 1, wherein the step S2 specifically comprises: Based on the architecture principle of the differentiable neural turing machine, constructing a space-time memory unit with an external memory matrix and a controller network; carrying out space channel coding processing on the standardized obstacle state input vector, extracting three-dimensional coordinate components of the standardized obstacle state input vector, mapping the three-dimensional coordinate components to a normalized grid space with preset resolution, generating a corresponding discretization grid index, and converting the discretization grid index into a position vector with fixed dimension by using an embedding layer; Based on the standardized obstacle state input vector extracted among continuous multiframes, calculating the relative displacement difference between adjacent time steps, and carrying out symbol quantization and interval division on the relative displacement difference to generate a dynamic mode label describing the change of the motion trend; The DNTM controller is utilized to receive the position vector and the dynamic mode label, double-channel joint coding operation is carried out, the position vector and the dynamic mode label are spliced and fused into a composite input vector, a hidden state is updated through GRU, a writing weight vector and a content vector are generated, a target address in a memory matrix is positioned according to the writing weight vector, the content vector is written into a corresponding position, and time-space coupling information storage action is completed; and repeatedly executing the double-channel joint coding operation, and continuously updating the memory matrix along with the step of simulation time to form an ordered memory feature matrix.
6. The simulation method for obstacle avoidance training of unmanned aerial vehicle inspection operation according to claim 1, wherein the step S3 specifically comprises: Based on a memory feature matrix stored by a space-time memory unit constructed by a differentiable neural turing machine structure, acquiring a space-time correlation sequence consisting of position vectors and relative displacement difference labels of each dynamic barrier in the last K time steps recorded in the memory feature matrix; Performing query-key-value mapping transformation on the space-time correlation sequence, generating a query vector by utilizing the environment observation state at the current moment, encoding historical track fragments into key vectors and value vectors, and calculating correlation scores between each historical fragment and the current evolution mode based on dot product attentiveness; Applying Softmax normalization processing to the relevance score to generate normalized relevance weight distribution; Based on the normalized correlation weight, performing weighted summation operation on the corresponding value vector sequence to generate a local evolution trend vector; and carrying out L2 normalization processing on the local evolution trend vector, and outputting dynamic behavior characterization of the barrier group.
7. The unmanned aerial vehicle inspection work obstacle avoidance training simulation method according to claim 6, wherein the local evolution trend vector explicitly codes the most influential obstacle movement trend in the current environment and the time-space dependency relationship thereof.
8. The simulation method for obstacle avoidance training of unmanned aerial vehicle inspection operation according to claim 1, wherein the step S4 specifically comprises: Constructing an obstacle node set based on a local evolution trend vector generated by weighted fusion of an attention mechanism, and taking the obstacle node set as an initial node characteristic of a graph structure; Establishing a non-directional edge to connect adjacent barrier nodes according to the relative distance threshold value of the barrier in the current three-dimensional space, and forming a dynamic adjacent matrix; Carrying out multi-layer neighborhood aggregation processing on the initial node characteristics and the dynamic adjacency matrix by using a graph convolution network, carrying out weighted summation on characteristic information of neighbor nodes by using a learnable weight parameter, extracting and updating high-order semantic characterization of each obstacle node, and obtaining an enhanced node embedded vector integrating context information; Based on the enhanced node embedded vector, calculating motion consistency measurement among obstacle nodes, identifying a space aggregation area with similar motion trend by adopting a clustering algorithm, judging the space aggregation area as a potential collision cluster, analyzing the overall displacement average value of the space aggregation area, and extracting a main stream motion direction vector; Combining the space expansion range of the potential collision cluster and the uncertainty variance of the main flow movement direction, modeling the position distribution in the future T steps by adopting a probability density estimation method, and outputting the space-time probability distribution envelope surface of each obstacle.
9. The unmanned aerial vehicle inspection work obstacle avoidance training simulation method of claim 8, wherein each node in the set of obstacle nodes characterizes a historical motion state sequence of a dynamic obstacle.
10. The simulation method for obstacle avoidance training of unmanned aerial vehicle inspection operation according to claim 1, wherein the step S5 specifically comprises: Based on an obstacle probability distribution envelope surface output by a graph convolution network, acquiring possible pose confidence intervals of each dynamic obstacle in future T steps, discretizing a continuous environment into a voxelized decision space with time slices by utilizing a three-dimensional space-time grid mapping algorithm in combination with the current position, pose and perceived visual field range of the unmanned aerial vehicle, and generating a local obstacle avoidance decision space body; In the local obstacle avoidance decision space body, based on unmanned aerial vehicle dynamic constraint and a maximum acceleration/angular velocity limit value, performing forward track sampling by adopting an improved RRT algorithm to generate N potential flight paths from the current pose, and forming a preliminary candidate flight path set; Performing collision risk assessment on space-time nodes on each candidate flight path, calculating accumulated collision cost of the paths crossing the high-confidence obstacle area by using the obstacle existence probability of the probability distribution envelope surface in the corresponding time slice, constructing a multidimensional cost function by combining the path length, curvature change and angle deviation from the task target direction, and outputting the comprehensive risk score of each path; non-dominant sorting and clustering pruning are carried out on the candidate flight path set based on the comprehensive risk score, P groups of path schemes with low risk and obvious diversity are reserved, each preferred path is mapped into a group of parameterized action strategy sequences, and a structured action strategy output set is generated; and carrying out space-time alignment binding on the action strategy output set and the corresponding local obstacle avoidance decision space feature map to form a pairing data tuple, and storing the pairing data tuple in a local buffer area in a standardized tensor format.

Description

Unmanned aerial vehicle inspection operation obstacle avoidance training simulation method Technical Field The invention relates to the technical field of real-time obstacle avoidance of dynamic obstacles in unmanned aerial vehicle inspection operation, in particular to an obstacle avoidance training simulation method for unmanned aerial vehicle inspection operation. Background Along with the wide application of unmanned aerial vehicle technology in complex environments such as power line inspection, forest monitoring, emergency rescue, the autonomous obstacle avoidance capability of the unmanned aerial vehicle technology has become a core technology foundation for guaranteeing safe and efficient operation. Currently, an unmanned aerial vehicle autonomous obstacle avoidance training and simulation system mainly depends on a physical simulation-based global environment reconstruction method or a depth time sequence network modeling dynamic obstacle behavior mode. Some mainstream schemes adopt traditional cyclic neural networks (such as RNN and LSTM) to predict the time sequence state of the obstacle, or combine with a high-computation-power physical engine to realize continuous simulation of the environment, so as to obtain a deduction result of the future state of the obstacle. In part of intelligent agent systems, the space relation is analyzed by utilizing a convolutional graph neural network, and the limited environment self-adaption capability is realized through a model migration strategy. The method is widely applied to the scenes such as airborne sensing, incremental strategy learning, high-fidelity simulation test and the like. However, the following disadvantages still exist in the prior art: (1) Dynamic obstacle behavior prediction accuracy is limited, and the current method is characterized by multiple single-step prediction or short-sequence historical modeling, so that nonlinear evolution process of complex obstacles in a dynamic environment is difficult to accurately describe. Particularly in scenes such as multi-obstacle real-time interaction, movement pattern mutation and the like, the future state prediction of the obstacle is easy to generate larger errors, and the effectiveness and the safety of the obstacle avoidance route of the unmanned aerial vehicle are directly affected. (2) The large-scale simulation based on the physical engine or the deep neural network requires strong calculation support, is limited by the real-time operation capability of the embedded platform and the edge equipment, is difficult to be efficiently deployed in the actual inspection operation scene of the unmanned aerial vehicle, is only suitable for batch training in experimental environments or cloud, and has limited real-time performance and portability. (3) The decision strategy migration efficiency is low, the existing knowledge migration method is mostly dependent on offline hard tag training or full model parameter synchronization, redundant information transmission is high, real-time mutation feedback of an environment mode is ignored, and therefore a student end network is difficult to quickly adapt to barrier evolution characteristics in a new environment, and autonomous decision capability improvement is limited. (4) The system lacks of an online self-adaptive mechanism with high robustness, faces sudden environmental changes or novel obstacle behavior modes, has low updating frequency of a traditional simulation system and a strategy model, cannot trigger timely model fine adjustment according to real-time observation residual errors, and is easy to cause potential safety hazards or decision delays. Disclosure of Invention The invention aims to solve the technical problems and provides an obstacle avoidance training simulation method for unmanned aerial vehicle inspection operation. The technical scheme of the invention is realized in such a way that the obstacle avoidance training simulation method for unmanned aerial vehicle inspection operation comprises the following steps: s1, acquiring a time sequence sample containing position, speed, acceleration and motion trend based on historical observation data of a dynamic obstacle in unmanned aerial vehicle inspection operation, and carrying out normalization processing on the sample to generate a standardized obstacle state input vector; s2, constructing a space-time memory unit by utilizing a differentiable neural turing machine structure, wherein based on the standardized obstacle state input vector, the position vector of a space channel and the relative displacement difference value label of a time channel are jointly encoded through a two-channel writing mechanism to form a memory feature matrix with space-time relevance; S3, reading obstacle state track segments of the last K time steps from the memory feature matrix, calculating the correlation weights of the segments and the current environment evolution mode by combining an attention mechanism