Urban logistics-oriented unmanned aerial vehicle adaptive anti-interference precise landing control system

Abstract

The invention discloses an unmanned aerial vehicle adaptive anti-interference precise landing control system for urban logistics, and the system comprises the following modules: a fractal intention tensor body construction module which is used for constructing a fractal intention tensor body; the asymmetric self-evolution intention generation network module is used for generating an intention section sequence; an anti-causal state projector module for generating a plurality of virtual future state nodes; the back projection function construction module is used for performing inverse mapping on each virtual future state node to generate a back-stepping control path set; the disturbance guide path search module is used for sampling and variation by adopting a disturbance guide Monte Carlo path search algorithm to generate an optimal control path; the anti-interference control instruction generation module is used for outputting an anti-interference control instruction; and the flight feedback updating module is used for collecting execution feedback information of the unmanned aerial vehicle and updating state evolution distribution of the fractal intention tensor body. According to the method, tensor modeling and disturbance optimization path backstepping are fused, and anti-interference accurate landing of the unmanned aerial vehicle is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle control, and particularly relates to an unmanned aerial vehicle adaptive anti-interference precision landing control system for urban logistics. BACKGROUND

[0002] With the rapid growth of urban low-altitude logistics and terminal unmanned distribution demand, the unmanned aerial vehicle autonomous landing and precision landing control technology in the complex urban environment of multi-source interference scene has attracted widespread attention. The existing urban unmanned aerial vehicle landing system mainly relies on GNSS / IMU fusion navigation, visual positioning or radar assisted landing path planning to adjust the attitude and terminal landing control, but in the dynamic disturbance environment, the following problems exist: The existing system has limited modeling capability for the spatial semantic information of the target landing area, often uses low-dimensional plane geometry or single-frame image features, lacks effective expression of platform three-dimensional structure, boundary topology and disturbance field multi-time and space evolution characteristics, and is difficult to support fine decision control; the existing control method mainly uses forward prediction path planning method, when affected by building reflection, signal shielding and electromagnetic interference nonlinear disturbance factors in the urban environment, the prediction accuracy and control stability decrease significantly, and it is difficult to realize the reverse reasoning and robust path planning of the future state; the heuristic algorithm based on graph search or fixed cost function is often used in path optimization process, which is difficult to integrate the dynamic feedback of the environment disturbance field, and cannot actively adapt to sudden interference or complex disturbance changes; in addition, the control strategy update usually relies on fixed period or threshold trigger mechanism, lacks real-time perception and iterative response to flight error trend and disturbance mode change, leading to high error accumulation risk and control response lag in the fine landing stage, and potential risks of aircraft instability or landing deviation.

[0003] Therefore, how to provide an unmanned aerial vehicle adaptive anti-interference precision landing control system for urban logistics is a problem that those skilled in the art need to solve. SUMMARY

[0004] One object of the present application is to provide an unmanned aerial vehicle adaptive anti-interference precision landing control system for urban logistics. The present application integrates multi-source fractal tensor modeling, asymmetric self-evolution intention generation network and disturbance guided Monte Carlo path search algorithm, constructs a control path system from the target state to the current state, realizes closed-loop adaptive control combined with flight feedback, has the advantages of strong anti-interference ability, low path energy consumption, high landing precision and adaptation to complex urban environment, and significantly improves the stability and control intelligence level of unmanned aerial vehicles in multi-source interference in urban low-altitude logistics.

[0005] The unmanned aerial vehicle adaptive anti-interference precision landing control system for urban logistics according to the embodiments of the present application comprises the following modules: a fractal intention tensor body construction module, configured to construct a fractal intention tensor body of a target landing state; an asymmetric self-evolution intention generation network module, configured to perform nonlinear coding and state evolution processing on the fractal intention tensor body to generate an intention section sequence; a reverse causality state projector module, including a target state pre-coding unit, a reverse reasoning path generation unit and a virtual future state node generation unit, configured to reversely deduce a reasoning path from the target landing state to a current unmanned aerial vehicle (UAV) state based on the intention section sequence, and generate a plurality of virtual future state nodes under interference modeling conditions; a reverse projection function construction module, configured to perform inverse mapping for each virtual future state node to generate a set of inverse push control paths starting from the current state; a perturbation guided path search module, configured to sample and mutate the set of inverse push control paths using a perturbation guided Monte Carlo path search algorithm, and select an inverse push control path with the minimum energy consumption value as an optimal control path; an anti-interference control instruction generation module, configured to generate and output an anti-interference control instruction based on a state error between the current state and the optimal control path and a perturbation condition; a flight feedback updating module, configured to collect UAV execution feedback information and update a state evolution distribution of the fractal intention tensor body in real time.

[0006] The UAV adaptive anti-interference precise landing control method for urban logistics according to the embodiment of the present application includes the following steps: Step one: constructing a fractal intention tensor body of a target landing state; Step two: performing nonlinear coding and evolution updating on the fractal intention tensor body using an asymmetric self-evolution intention generation network to generate an intention section sequence; Step three: based on the intention section sequence, calling a reverse causality state projector module to construct a reasoning path reversely deduced from the target landing state to the current UAV state, and generating a plurality of virtual future state nodes under interference modeling conditions; Step four: constructing a reverse projection function for each virtual future state node to inversely map the virtual future state to the current state to form a set of inverse push control paths; Step five: sampling and mutating the set of inverse push control paths using a perturbation guided Monte Carlo path search algorithm, and selecting an inverse push control path with the minimum energy consumption value as an optimal control path; Step six: generating an anti-interference control instruction based on the optimal control path to drive the UAV to perform anti-interference flight control; Step seven: according to the feedback information of the unmanned aerial vehicle after executing the anti-interference control instruction, the state evolution distribution of the fractal intention tensor body is updated in real time, and steps two to six are iteratively executed until the unmanned aerial vehicle completes the precise landing operation on the target landing platform.

[0007] Optionally, the fractal intention tensor body is a four-dimensional tensor , wherein represents a transverse spatial coordinate, represents a longitudinal spatial coordinate, represents a vertical spatial coordinate, represents a time step relative to the starting time of the flight task, and the construction process of the fractal intention tensor body includes: acquiring three-dimensional geometric structure data of the target landing platform, using point cloud data acquired by a laser radar scanning system, and forming a first sub-tensor through voxelization processing, for describing the spatial structure of the target landing platform, including platform outer boundary, landing surface normal variation, and three-dimensional concave-convex features; processing real-time image frames based on a convolutional neural network image semantic segmentation method, identifying platform edges, landing mark lines, and fixed obstacle edges in the target landing platform region, encoding into a two-dimensional semantic classification matrix, and constructing a second sub-tensor ; collecting three-dimensional wind speed vector data of a wind speed sensor, angular velocity and linear acceleration data of an inertial measurement sensor, data of an electromagnetic interference intensity sensor, and positioning deviation data, respectively, and reconstructing through time alignment and spatial interpolation to construct a third sub-tensor , for reflecting the dynamic distribution characteristics of disturbances at different spatial positions and time points; dimensionally standardizing and processing the first sub-tensor, the second sub-tensor, and the third sub-tensor, and splicing the tensor channels to form a complete four-dimensional fractal intention tensor body.

[0008] Optionally, the asymmetric self-evolution intention generation network specifically includes: an input preprocessing layer for tensor normalization and time step block processing of the input fractal intention tensor body, decomposing the fractal intention tensor body into a series of time slice tensors; an asymmetric multi-scale feature encoder including multiple parallel convolution branches, each convolution branch having different convolution kernel sizes, step lengths, and channel depths, and including at least the following three groups of configurations: a first convolution branch using a kernel size of with a step length of 1, for capturing local spatial structure features; a second convolution branch using a kernel size of with a step length of 2, for extracting mesoscale disturbance pattern changes; The third convolution branch uses a kernel size of , a step size of 1, and only acts on the horizontal direction to capture the semantic expansion structure of the platform boundary; The feature maps output by each convolution branch are fused into a unified encoding feature tensor through channel concatenation; The time evolution self-attention module groups the unified encoding feature tensor into a sequence according to the time steps and inputs it into a multi-head self-attention structure to generate a time sequence feature; The gating graph structure evolution recurrent unit takes the time sequence feature as the initial graph node state, constructs a dynamic evolution graph structure, with each time step corresponding to a graph node, builds graph edges based on spatial adjacency relationships and historical disturbance distribution, and introduces a gating mechanism to adjust the edge weight update process over time. The initial value of the edge weight is obtained by weighted calculation based on the consistency of the three-dimensional Euclidean distance between nodes and the gradient direction of the disturbance field. The output is a sequence of encoded potential intent states; The intent section reconstruction layer includes a convolution upsampling module and a residual correction branch, which receives the potential intent state sequence as input, sequentially restores it to the original spatial dimension through the convolution upsampling module, and performs detail compensation through the residual correction branch, and outputs a time-ordered intent section sequence.

[0009] Optionally, the anti-causal state projector module specifically includes: A target state pre-encoding unit inputs the intent section sequence, encodes the spatial structure information and semantic features in the target landing state using a fully connected network and a position embedding method to generate a target state embedding vector; A reverse reasoning path generation unit receives the target state embedding vector as input and gradually reverses to the current UAV state through a reverse recurrent neural network unit, including: setting the target state embedding vector as the initial input state, executing the state update operation of the reverse recurrent unit in reverse time order, and at each step, the potential position, attitude and velocity of the UAV are reversely predicted according to the hidden state vector at the previous time, until the initial state of the current UAV, forming a continuous reverse reasoning path from the target landing state to the current UAV state; A virtual future state node generation unit receives the reverse reasoning path and generates multiple virtual future state nodes through disturbance field simulation and random sampling methods under disturbance modeling conditions, including: Wind field disturbance condition: randomly generate disturbance wind speed and direction in the reverse reasoning path space to simulate wind-induced disturbance field; Global navigation satellite system disturbance condition: randomly generate GNSS signal errors or interruptions in a specified area to simulate GNSS disturbance field; Electromagnetic interference condition: randomly inject electromagnetic interference intensity change nodes in the reverse reasoning path to simulate the electromagnetic disturbance field of urban space; By the interference modeling condition, the reverse reasoning path is disturbed, disturbed and randomly varied to obtain a plurality of virtual future state nodes containing interference disturbance information.

[0010] Optionally, the step four: constructing a back projection function for each virtual future state node, inversely mapping the virtual future state to the current state to form a set of inverse control paths, specifically: For each virtual future state node, a back projection function is established, and the back projection function is realized by a neural network structure with attention mechanism, including an input layer, a back projection coding layer, an attention mapping layer and an output decoding layer; The input layer is used to receive the spatial coordinates, flight attitude, speed and corresponding disturbance field information of the virtual future state node; The back projection coding layer uses a multi-layer perception machine to encode the input virtual future state to obtain a state embedding vector; The attention mapping layer includes a multi-head attention structure, takes the initial state feature vector of the current UAV state as the query vector, and takes the state embedding vector output by the back projection coding layer as the key and value vectors. The inverse mapping relationship between the current state and the virtual future state is calculated through the attention mechanism; The output decoding layer decodes the output result of the attention mapping layer into a specific flight control action sequence; The back projection function is used to inversely map each virtual future state node step by step, and the inverse mapping process is performed in reverse time sequence order. Each step of inverse mapping inversely calculates the potential control state of the previous step according to the features and disturbance field conditions of the next virtual future state, until the inverse mapping returns to the current state; The inverse mapping process is performed for each virtual future state node to obtain a control action sequence from the virtual future state node back to the current state; The control action sequence is arranged one by one according to the corresponding virtual future state node to form a set of inverse control paths starting from the current state.

[0011] Optionally, the step five: for the set of inverse control paths, a disturbance guided Monte Carlo path search algorithm is used for sampling and variation, and the inverse control path with the minimum energy consumption value is selected as the optimal control path, specifically: For each inverse control path, a disturbance sampling space is constructed according to the spatial distribution position, and a plurality of rounds of path variation operations are performed in the disturbance sampling space by using the Monte Carlo path search algorithm: Spatial coordinate perturbations are injected at random nodes in the inverse control path. The perturbation amplitude is limited according to the gradient change of the corresponding perturbation source to ensure that the perturbation direction is consistent with the perturbation vector field direction. Curvature perturbations are introduced between path segments, and the curvature radius of the connection points is modified to simulate the trajectory deviation effect of the aircraft under crosswind perturbation. A time delay perturbation is added to a set path segment, and the trajectory drift caused by the thrust response delay is simulated by extending the time step of the set path segment; Calculate the flight distance, total attitude adjustment angle, and control force input integral of the path segment in the disturbance region for each reverse thrust control path, and then sum them up by weight to obtain the energy consumption value of the reverse thrust control path; From all evaluated back-engineering control paths, the path with the lowest energy consumption is selected as the optimal control path.

[0012] Optionally, step six: Based on the optimal control path, generating anti-interference control commands to drive the UAV to perform anti-interference flight control, specifically: Let the current state of the drone be... ,in Represents the current three-dimensional spatial position vector. Represents the current velocity vector. Represents the current attitude angle vector; Let the target state at the corresponding time in the optimal control path be... ,in The three-dimensional spatial position vector representing the optimal control path. The velocity vector represents the optimal control path. The attitude angle vector representing the optimal control path; Calculate the state error, including: Position error , representing the Euclidean distance between the current position and the target position; speed error , representing the magnitude difference between the current velocity and the target velocity; Attitude error , representing the angle difference between the target attitude and the current attitude; And obtain the disturbance intensity of the current flight environment. ; Preset position error tolerance Speed ​​error tolerance Attitude error tolerance and the upper limit threshold of disturbance intensity It generates anti-interference control commands based on the set control command discrimination rules.

[0013] Optionally, the setting control instruction discrimination rule is specifically: If , , and are met, the basic control instruction is output, and the attitude and thrust are adjusted according to the optimal control path; If any state error exceeds the corresponding tolerance or the disturbance intensity , the anti-disturbance compensation control mechanism is triggered, the control parameter is dynamically adjusted, and the feedforward disturbance compensation term is added; If , and all exceed the set upper limit in continuous multiple control periods, and are higher than for a long time, the protection mode is entered, and the response sensitivity of the aircraft is limited to prevent flight instability; The finally generated anti-interference control instruction includes the attitude angle adjustment amount, the thrust vector distribution value and the time synchronization control parameter, and is sent to the flight control system to drive the unmanned aerial vehicle to execute the anti-interference flight according to the optimal control path.

[0014] Optionally, the state evolution distribution of the fractal intention tensor body is updated in real time according to the feedback information of the unmanned aerial vehicle after executing the anti-interference control instruction, and the specific process is as follows: In each control period, the actual flight feedback information of the unmanned aerial vehicle is collected, including the current position, speed, attitude angle and acceleration state data, and the feedback information is compared with the target state in the optimal control path in time sequence, to obtain the deviation change sequence between the actual state and the target state; According to the deviation change sequence, the state error evolution trend is constructed, and combined with the environmental disturbance modeling information of the current region, the error evolution trend is fused and coded with the disturbance mode as a dynamic correction signal input into the asymmetric self-evolution intention generation network; In the fractal intention tensor body, the time slice tensor corresponding to the current control period is located and replaced with the dynamic correction signal, so that the local intention update at the current time is completed.

[0015] The beneficial effects of the present application are: The application constructs a fractal intention tensor body containing the target landing platform geometry, spatial boundary semantic information and dynamic distribution of the disturbance field, carries out multi-scale nonlinear coding and time sequence evolution modeling on the tensor body by combining an asymmetric self-evolution intention generation network, proposes a reverse path construction method based on anti-causal state reasoning for the path uncertainty and control lag problem under multi-source non-Gaussian disturbance in a complex urban environment, generates a virtual future state node under interference through reverse cyclic reasoning, realizes inverse mapping backtracking of high-dimensional state by using an attention-guided back projection function, obtains a refined reverse control path set, introduces a disturbance-guided Monte Carlo path search algorithm in the path optimization process, combines spatial disturbance sampling and multi-round curvature / delay variation mechanism, accurately evaluates the path energy consumption and selects the energy-optimal control path under global disturbance, and fuses the attitude, position, speed error and disturbance intensity in the control instruction generation stage, establishes a dynamic discrimination rule to generate a robust anti-interference control instruction, and based on flight feedback information, the tensor body state distribution is corrected in real time to realize closed-loop adaptive update of the control strategy, the system significantly improves the landing accuracy, flight stability and control response speed of the unmanned aerial vehicle in the multi-source disturbance environment in the urban low-altitude logistics scene, and has good generalization ability and engineering deployability. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application. In the drawings: Figure 1 The structure diagram of the unmanned aerial vehicle adaptive anti-interference precise landing control system for urban logistics proposed by the application; Figure 2 The overall flowchart of the unmanned aerial vehicle adaptive anti-interference precise landing control method for urban logistics proposed by the application. DETAILED DESCRIPTION

[0017] The application will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0018] Reference Figure 1 The unmanned aerial vehicle adaptive anti-interference precise landing control system for urban logistics comprises the following modules: A fractal intention tensor body construction module is used to construct a fractal intention tensor body of the target landing state; An asymmetric self-evolution intention generation network module is used to perform nonlinear coding and state evolution processing on the fractal intention tensor body, and generate an intention section sequence; The anti-causal state projector module includes a target state pre-encoding unit, a reverse reasoning path generation unit, and a virtual future state node generation unit, configured to deduce a reasoning path from the target landing state to the current UAV state in reverse based on the intention section sequence, and generate a plurality of virtual future state nodes under interference modeling conditions; The anti-projection function construction module is configured to perform inverse mapping for each virtual future state node to generate a set of inverse push control paths starting from the current state; The disturbance guided path search module is configured to sample and mutate the set of inverse push control paths using a disturbance guided Monte Carlo path search algorithm, and select the inverse push control path with the minimum energy consumption value as the optimal control path; The anti-interference control instruction generation module is configured to generate and output an anti-interference control instruction based on the state error between the current state and the optimal control path and the disturbance condition; The flight feedback update module is configured to collect UAV execution feedback information and update the state evolution distribution of the fractal intention tensor body in real time.

[0019] Reference Figure 2 The UAV adaptive anti-interference precision landing control method for urban logistics comprises the following steps: Step one: build a fractal intention tensor body of the target landing state; Step two: use the asymmetric self-evolution intention generation network to perform nonlinear encoding and evolution update on the fractal intention tensor body to generate an intention section sequence; Step three: based on the intention section sequence, call the anti-causal state projector module to build a reasoning path from the target landing state to the current UAV state in reverse, and generate a plurality of virtual future state nodes under interference modeling conditions; Step four: construct an anti-projection function for each virtual future state node to inversely map the virtual future state to the current state to form a set of inverse push control paths; Step five: for the set of inverse push control paths, use a disturbance guided Monte Carlo path search algorithm to sample and mutate, and select the inverse push control path with the minimum energy consumption value as the optimal control path; Step six: based on the optimal control path, generate an anti-interference control instruction to drive the UAV to perform anti-interference flight control; Step seven: based on the feedback information after the UAV executes the anti-interference control instruction, update the state evolution distribution of the fractal intention tensor body in real time, and iteratively execute steps two to six until the UAV completes the precision landing operation on the target landing platform.

[0020] In this embodiment, the fractal intention tensor body is a four-dimensional tensor wherein Represents horizontal spatial coordinates. Represents the vertical spatial coordinates. Represents vertical spatial coordinates. The construction process of the fractal intent tensor, representing the time step relative to the start time of the flight mission, includes: The three-dimensional geometric structure data of the target landing platform is acquired using point cloud data obtained from a lidar scanning system, which is then processed into a voxelization model to form the first subtensor. It is used to describe the spatial structure of the target landing platform, including the platform's outer boundary, the normal variation of the take-off and landing surface, and its three-dimensional concave and convex features; Real-time image frames are processed using an image semantic segmentation method based on convolutional neural networks. Platform edges, landing marker lines, and fixed obstacle edges in the target landing platform area are identified and encoded into a two-dimensional semantic classification matrix, which is then used to construct a second sub-tensor. ; Three-dimensional wind speed vector data from an anemometer, angular velocity and linear acceleration data from an inertial measurement sensor, data from an electromagnetic interference intensity sensor, and positioning deviation data were collected respectively. A third subtensor was then constructed through time alignment and spatial interpolation reconstruction. It is used to reflect the dynamic distribution characteristics of disturbances at different spatial locations and time points; The first, second, and third subtensors are dimensionally normalized and spliced ​​with tensor channels to form a complete four-dimensional fractal intent tensor volume.

[0021] In this embodiment, the asymmetric self-evolutionary intent generation network specifically includes: The input preprocessing layer is used to perform tensor normalization and time step block processing on the input fractal intent tensor volume, decomposing the fractal intent tensor volume into a series of time slice tensors. An asymmetric multi-scale feature encoder comprises multiple parallel convolutional branches, each with different kernel size, stride, and channel depth, and includes at least the following three configurations: The first convolutional branch uses a kernel size of... The step size is 1, which is used to capture local spatial structure features; The second convolutional branch uses a kernel size of... The step size is 2, which is used to extract mesoscale perturbation pattern changes; The third convolutional branch uses a kernel size of... With a step size of 1, it operates only in the horizontal direction and is used to capture the semantic extension structure of the platform boundary; The feature maps output by each convolutional branch are fused into a unified encoded feature tensor by channel concatenation. A time evolution self-attention module groups the unified encoded feature tensors by time steps to form a sequence and inputs the sequence into a multi-head self-attention structure to generate time sequence features; A gating graph structure evolution loop unit takes the time sequence features as initial graph node states, constructs a dynamic evolution graph structure, each time step corresponding to a graph node, constructs graph edges based on spatial adjacency relationships and historical disturbance distributions, and introduces a gating mechanism to adjust the updating process of edge weights over time. The initial value of the edge weight is obtained based on the weighted calculation of the consistency between the three-dimensional Euclidean distance between nodes and the gradient direction of the disturbance field. The output is a sequence of encoded potential intent states. An intent cross-section reconstruction layer includes a convolutional upsampling module and a residual correction branch, which receives the sequence of potential intent states as input, sequentially restores to the original spatial dimension through the convolutional upsampling module, and performs detail compensation through the residual correction branch, and outputs a time-ordered sequence of intent cross-sections.

[0022] In this embodiment, the anti-causal state projector module specifically includes: A target state pre-encoding unit inputs the intent cross-section sequence, encodes the spatial structure information and semantic features in the target landing state using a fully connected network and a position embedding method to generate a target state embedding vector. A reverse reasoning path generation unit receives the target state embedding vector as input, gradually reverses to the current UAV state through a reverse recurrent neural network unit, including: setting the target state embedding vector as the initial input state, sequentially executing the state update operation of the reverse recurrent unit in reverse time order, and at each step, the potential position, attitude and velocity of the UAV are reversely predicted according to the hidden state vector at the previous time, until the initial state of the current UAV, forming a continuous reverse reasoning path from the target landing state to the current UAV state. A virtual future state node generation unit receives the reverse reasoning path and generates multiple virtual future state nodes through disturbance field simulation and random sampling methods under disturbance modeling conditions, including: Wind field disturbance condition: randomly generate disturbance wind speed and wind direction in the reverse reasoning path space to simulate wind-induced disturbance field; Global navigation satellite system disturbance condition: randomly generate GNSS signal errors or interruptions in a specified area to simulate GNSS disturbance field; Electromagnetic interference condition: randomly inject electromagnetic interference intensity change nodes in the reverse reasoning path to simulate the electromagnetic disturbance field in urban space; The reverse reasoning path is disturbed and disturbed by the disturbance modeling conditions to obtain multiple virtual future state nodes containing disturbance information.

[0023] In this embodiment, step four: for each virtual future state node, a back projection function is constructed to inversely map the virtual future state to the current state, forming a set of inverse control paths, specifically: For each virtual future state node, a back projection function is established, which is realized by a neural network structure with attention mechanism, including an input layer, a back projection encoding layer, an attention mapping layer, and an output decoding layer. The input layer is used to receive the spatial coordinates, flight attitude, speed, and corresponding disturbance field information of the virtual future state node. The back projection encoding layer uses a multi-layer perceptron to encode the input virtual future state to obtain a state embedding vector. The attention mapping layer includes a multi-head attention structure, taking the initial state feature vector of the current UAV state as the query vector, and the state embedding vector output by the back projection encoding layer as the key and value vectors. The inverse mapping relationship between the current state and the virtual future state is calculated through the attention mechanism. The output decoding layer decodes the output of the attention mapping layer into a specific flight control action sequence. The back projection function is used to perform step-by-step inverse mapping for each virtual future state node. The inverse mapping process is performed in reverse time sequence, and each step of inverse mapping is based on the features and disturbance field conditions of the next virtual future state to calculate the potential control state of the previous step, until the inverse mapping returns to the current state. For each virtual future state node, the inverse mapping process is performed to obtain the control action sequence from the virtual future state node back to the current state. The control action sequence is arranged according to the corresponding virtual future state node to form a set of inverse control paths starting from the current state.

[0024] In this embodiment, step five: for the set of inverse control paths, a disturbance-guided Monte Carlo path search algorithm is used for sampling and mutation, and the inverse control path with the smallest energy consumption value is selected as the optimal control path, specifically: For each inverse control path, a disturbance sampling space is constructed according to the spatial distribution position, and the Monte Carlo path search algorithm is used to perform multiple rounds of path mutation operations in the disturbance sampling space: A spatial coordinate disturbance is injected at a random node in the inverse control path, and the disturbance amplitude is limited according to the gradient change of the corresponding disturbance source to ensure that the disturbance direction is consistent with the disturbance vector field direction. Curvature disturbance is introduced between path segments to modify the curvature radius of the connection point, simulating the trajectory deviation effect of the aircraft under crosswind disturbance. The time delay disturbance is added to the set path segment, and the trajectory drift caused by the thrust response delay is simulated by extending the time step of the set path segment; The flight distance, attitude adjustment angle sum, and disturbance region path segment control force input integral corresponding to each backstepping control path are calculated, and the energy consumption values of the backstepping control paths are obtained by weighted summation. From all the evaluated backstepping control paths, the path with the minimum energy consumption value is selected as the optimal control path.

[0025] In the embodiment, the step six: based on the optimal control path, an anti-interference control instruction is generated to drive the UAV to perform anti-interference flight control, specifically: Let the current UAV state be , wherein represents the current three-dimensional space position vector, represents the current velocity vector, represents the current attitude angle vector, including the pitch angle, yaw angle, and roll angle; Let the target state corresponding to the time in the optimal control path be , wherein represents the three-dimensional space position vector of the optimal control path, represents the velocity vector of the optimal control path, represents the attitude angle vector of the optimal control path, including the pitch angle, yaw angle, and roll angle; The state error is calculated, including: The position error , which represents the Euclidean distance between the current position and the target position; The velocity error , which represents the difference in module length between the current velocity and the target velocity; The attitude error , which represents the angle difference between the target attitude and the current attitude; And the disturbance intensity of the current flight environment is obtained; The position error tolerance , the velocity error tolerance , the attitude error tolerance , and the upper limit threshold of the disturbance intensity are preset, and the anti-interference control instruction is generated according to the set control instruction discrimination rule.

[0026] In the embodiment, the set control instruction discrimination rule is specifically: If , , and Then, basic control commands are output to adjust attitude and thrust according to the optimal control path; If any state error exceeds the corresponding tolerance or disturbance strength If so, the disturbance rejection compensation control mechanism is triggered, the control parameters are dynamically adjusted and a feedforward disturbance compensation term is added; If within multiple consecutive control cycles , and All exceeded the set limit, and Long duration above If this occurs, the aircraft will enter protection mode, limiting its response sensitivity to prevent flight instability. The final anti-interference control commands include attitude angle adjustment, thrust vector allocation, and time synchronization control parameters, and are sent to the flight control system to drive the UAV to perform anti-interference flight according to the optimal control path.

[0027] In this embodiment, the step of updating the state evolution distribution of the fractal intent tensor in real time based on the feedback information after the UAV executes the anti-interference control command specifically involves: In each control cycle, the actual flight feedback information of the UAV is collected, including current position, speed, attitude angle and acceleration status data, and the feedback information is compared with the target state in the optimal control path in time sequence to obtain the deviation change sequence between the actual state and the target state. The state error evolution trend is constructed based on the deviation change sequence, and combined with the environmental disturbance modeling information of the current region, the error evolution trend and disturbance mode are fused and encoded, and input as a dynamic correction signal into the asymmetric self-evolution intention generation network. In the fractal intent tensor, the time slice tensor corresponding to the current control cycle is located and replaced with the dynamic correction signal, thereby completing the local intent update at the current moment.

[0028] Example 1

[0029] To verify the feasibility of this invention in practice, it was applied to an automated landing system for logistics drones in a complex urban environment. The target scenario was a precision delivery platform for drones between high-rise buildings, where multiple sources of disturbance existed, including strong wind disturbances, electromagnetic signal blockage, GNSS signal drift, and local obstacle obstruction. Traditional landing control methods based on global positioning or inertial navigation suffer from significant drawbacks in such environments, such as large landing deviations, lag in attitude adjustment, and slow anti-disturbance response. These drawbacks can easily lead to landing failures or damage to the drone, severely impacting the actual deployment efficiency and safety of urban logistics drones.

[0030] In actual tests, a group of six-rotor logistics unmanned aerial vehicles equipped with the control system of the application are selected, and five typical landing test platforms are set up, i.e., a roof between buildings, a narrow balcony, a roof with obstacles, a steel structure platform, and an internal vertical shaft with height restrictions. The tests are completed by using the system of the application to control the whole process from entering the interference field to the precise landing point. In this process, the wind speed, electromagnetic intensity, GNSS signal stability, flight attitude, and control response parameters are collected in real time for each flight task, and the method of the application is compared with the traditional PID control method.

[0031] In this embodiment, the system of the application enhances the perception accuracy of the spatial semantics and disturbance state of the target platform by constructing a fractal intention tensor body containing structure information, semantic boundaries, and disturbance distribution, predicts the intention evolution sequence at multiple times by using an asymmetric self-evolution intention generation network, constructs an anti-causal state projector to reversely deduce the dynamic state path between the current state of the unmanned aerial vehicle and the target landing state, and constructs an anti-projection function for the virtual future state, combined with the disturbance guided Monte Carlo path search algorithm, to select the energy optimal reverse path in multiple rounds of mutation-selection, thereby generating a high-robustness, dynamically adjustable control instruction, and finally realizing stable, precise, and safe anti-interference control.

[0032] Table 1 Comparison data table of the method of the application and the traditional PID control method

[0033] As can be seen from the data in Table 1, the unmanned aerial vehicle using the system of the application still maintains a maximum position error of not more than 0.35 m, an average attitude error of within 4.8°, and a control system response delay of less than 85 ms in a high wind speed (maximum wind speed 8.1 m / s) environment; while the unmanned aerial vehicle using the traditional PID control system has a maximum deviation of 1.3 m, an attitude deviation of more than 15°, and a slight collision occurs in two tests under the same conditions.

[0034] This embodiment verifies the robustness advantage of the system of the application in the background of multiple interference superposition, and also reflects the high adaptability and dynamic adjustment capability of the fractal intention tensor body and the disturbance guided Monte Carlo path search algorithm introduced in the system to the flight control path generation. The system effectively offsets the cumulative error caused by the disturbance through continuous feedback correction and path iteration optimization, ensuring that the aircraft completes precise landing with the minimum energy cost. Therefore, the application has significant engineering popularization value in the application of unmanned aerial vehicles with high precision and high interference requirements such as urban logistics.

[0035] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A UAV adaptive anti-interference precision landing control system for urban logistics, characterized in that, The method comprises the following modules: a fractal intention tensor body construction module, configured to construct a fractal intention tensor body of a target landing state; an asymmetric self-evolution intention generation network module, configured to perform nonlinear coding and state evolution processing on the fractal intention tensor body to generate an intention section sequence; a reverse causality state projector module, comprising a target state pre-coding unit, a reverse reasoning path generation unit, and a virtual future state node generation unit, configured to deduce a reasoning path from the target landing state to a current UAV state in reverse based on the intention section sequence, and generate a plurality of virtual future state nodes under interference modeling conditions; a reverse projection function construction module, configured to perform inverse mapping for each virtual future state node to generate a set of inverse push control paths from the current state; a perturbation guided path search module, configured to sample and mutate the set of inverse push control paths using a perturbation guided Monte Carlo path search algorithm, and select an inverse push control path with the minimum energy consumption value as an optimal control path; an anti-interference control instruction generation module, configured to generate and output an anti-interference control instruction based on a state error between the current state and the optimal control path and a perturbation condition; a flight feedback updating module, configured to collect UAV execution feedback information and update a state evolution distribution of the fractal intention tensor body in real time. 2.The urban logistics oriented UAV adaptive anti-interference precision landing control system of claim 1, wherein, The modules are connected through the following methods: Step 1: Construct a fractal intention tensor body of a target landing state; Step 2: Perform nonlinear coding and evolution updating on the fractal intention tensor body using an asymmetric self-evolution intention generation network to generate an intention section sequence; Step 3: Based on the intention section sequence, call the reverse causality state projector module to construct a reasoning path deduced from the target landing state to the current UAV state in reverse, and generate a plurality of virtual future state nodes under interference modeling conditions; Step 4: For each virtual future state node, construct a reverse projection function to inversely map the virtual future state to the current state to form a set of inverse push control paths; Step 5: For the set of inverse push control paths, sample and mutate using a perturbation guided Monte Carlo path search algorithm, and select an inverse push control path with the minimum energy consumption value as an optimal control path; Step 6: Based on the optimal control path, generate an anti-interference control instruction to drive the UAV to perform anti-interference flight control; Step 7: Based on the feedback information after the UAV executes the anti-interference control instruction, update the state evolution distribution of the fractal intention tensor body in real time, and iteratively execute steps 2 to 6 until the UAV completes the precise landing operation on the target landing platform. 3.The urban logistics oriented UAV adaptive anti-interference precision landing control system according to claim 2, characterized in that, The fractal intention tensor body is a four-dimensional tensor, and the construction process of the fractal intention tensor body comprises: acquiring three-dimensional geometric structure data of a target landing platform, using point cloud data acquired by a laser radar scanning system, and performing voxelization processing to form a first sub-tensor; processing real-time image frames based on a convolutional neural network image semantic segmentation method, identifying platform edges, landing marker lines, and fixed obstacle edges in the target landing platform region, encoding into a two-dimensional semantic classification matrix, and constructing a second sub-tensor; The three-dimensional wind speed vector data of the wind speed sensor, the angular velocity and linear acceleration data of the inertial measurement sensor, the data of the electromagnetic interference intensity sensor, and the positioning deviation data are collected respectively, and are reconstructed through time alignment and spatial interpolation to construct a third sub-tensor; The first sub-tensor, the second sub-tensor and the third sub-tensor are subjected to dimension standardization processing and tensor channel splicing to form a complete four-dimensional fractal intention tensor body. 4.The urban logistics oriented UAV adaptive anti-interference precision landing control system of claim 2, wherein, The asymmetric self-evolution intention generation network specifically comprises: An input preprocessing layer is configured to perform tensor normalization and time step blocking processing on the input fractal intention tensor body, and decompose the fractal intention tensor body into a series of time slice tensors; An asymmetric multi-scale feature encoder comprises a plurality of parallel convolution branches, each convolution branch having different convolution kernel sizes, steps and channel depths, and at least comprising the following three groups of configurations: The first convolution branch uses a kernel size of with a step size of 1 to capture local spatial structure features; The second convolution branch uses a kernel size of with a step size of 2 for extracting mesoscale perturbation pattern changes; The third convolution branch uses a kernel size of , a step of 1, and only acts on the horizontal direction to capture the semantic expansion structure of the platform boundary; The feature maps output by each convolution branch are fused into a unified encoding feature tensor through channel splicing; A time evolution self-attention module groups the unified encoding feature tensor into a sequence according to time steps, and inputs the sequence into a multi-head self-attention structure to generate a time sequence feature; A gated graph structure evolution recurrent unit takes the time sequence feature as an initial graph node state, constructs a dynamic evolution graph structure, each time step corresponding to a graph node, constructs graph edges based on spatial adjacency relationships and historical disturbance distribution, and introduces a gating mechanism to adjust the updating process of edge weights over time. The initial value of the edge weight is obtained based on the consistency of the three-dimensional Euclidean distance between nodes and the gradient direction of the disturbance field. An encoded latent intention state sequence is output. An intention section reconstruction layer comprises a convolution up-sampling module and a residual correction branch, which receives the latent intention state sequence as input, sequentially restores to the original spatial dimension through the convolution up-sampling module, and performs detail compensation through the residual correction branch, and outputs an intention section sequence arranged in time sequence. 5.The urban logistics oriented UAV adaptive anti-interference precision landing control system of claim 2, wherein, The anti-causal state projector module specifically comprises: A target state pre-encoding unit inputs the intention section sequence, encodes the spatial structure information and semantic features in the target landing state using a fully connected network and a position embedding method to generate a target state embedding vector; A reverse reasoning path generation unit receives the target state embedding vector as input, and gradually reverses to the current unmanned aerial vehicle state through a reverse recurrent neural network unit, including: setting the target state embedding vector as the initial input state, sequentially executing the state update operation of the reverse recurrent unit in reverse time sequence, and at each step, the potential position, attitude and speed of the unmanned aerial vehicle are reversely predicted according to the hidden state vector at the previous time, until the initial state of the current unmanned aerial vehicle, forming a continuous reverse reasoning path from the target landing state to the current unmanned aerial vehicle state; A virtual future state node generation unit receives the reverse reasoning path, generates a plurality of virtual future state nodes through disturbance field simulation and random sampling method under the interference modeling condition, and the interference modeling condition comprises: Wind field interference condition: randomly generate disturbance wind speed and wind direction in the reverse reasoning path space to simulate wind-induced disturbance field; Global Navigation Satellite System interference condition: randomly generate GNSS signal error or interruption in a set area to simulate GNSS interference field; Electromagnetic interference condition: randomly inject electromagnetic interference intensity change nodes in the backward reasoning path to simulate the electromagnetic disturbance field in urban space; The interference modeling condition is used to disturb and disrupt the backward reasoning path, and a plurality of virtual future state nodes containing interference disturbance information are obtained. 6.The urban logistics oriented UAV adaptive anti-interference precision landing control system of claim 2, wherein, Step four: For each virtual future state node, a back projection function is constructed to inversely map the virtual future state to the current state, forming a set of inverse control paths, specifically: For each virtual future state node, a back projection function is established, which is realized by a neural network structure with attention mechanism, including an input layer, a back projection encoding layer, an attention mapping layer, and an output decoding layer; The input layer is used to receive the spatial coordinates, flight attitude, speed of the virtual future state node, and the corresponding disturbance field information; The back projection encoding layer uses a multi-layer perception machine to encode the input virtual future state to obtain a state embedding vector; The attention mapping layer includes a multi-head attention structure, taking the initial state feature vector of the current UAV state as the query vector, and the state embedding vector output by the back projection encoding layer as the key and value vectors, and calculating the inverse mapping relationship between the current state and the virtual future state through the attention mechanism; The output decoding layer decodes the output result of the attention mapping layer into a specific flight control action sequence; The back projection function is used to inversely map each virtual future state node step by step, and the inverse mapping process is performed in reverse time sequence order. Each step of inverse mapping is based on the features and disturbance field conditions of the next virtual future state to inversely calculate the potential control state of the previous step, until the inverse mapping is back to the current state; For each virtual future state node, the inverse mapping process is performed to obtain the control action sequence from the virtual future state node back to the current state; The control action sequences are arranged one by one according to the corresponding virtual future state nodes to form a set of inverse control paths starting from the current state. 7.The urban logistics oriented UAV adaptive anti-interference precision landing control system according to claim 2, characterized in that, Step five: For the set of inverse control paths, a disturbance-guided Monte Carlo path search algorithm is used for sampling and variation, and the inverse control path with the smallest energy consumption value is selected as the optimal control path, specifically: For each inverse control path, a disturbance sampling space is constructed according to the spatial distribution position, and a multi-round path variation operation is performed in the disturbance sampling space using the Monte Carlo path search algorithm: Inject spatial coordinate disturbance at random nodes in the inverse control path, and limit the disturbance amplitude according to the gradient change of the corresponding disturbance source to ensure that the disturbance direction is consistent with the disturbance vector field direction; Introduce curvature disturbance between path segments to modify the curvature radius of the connection point to simulate the trajectory deviation effect of the aircraft under crosswind disturbance; Add time delay disturbance to the set path segment to simulate the trajectory drift caused by thrust response delay by extending the time step of the set path segment. The flight distance corresponding to each backstepping control path, the total attitude adjustment angle, and the control force input integral of the disturbance region path segment are calculated, and the energy consumption values of the backstepping control paths are obtained by weighted summation. From all the evaluated backstepping control paths, the path with the minimum energy consumption value is selected as the optimal control path. 8.The urban logistics oriented UAV adaptive anti-interference precision landing control system of claim 2, wherein, The step six: based on the optimal control path, generate anti-interference control instructions to drive the UAV to perform anti-interference flight control, specifically: Let the current state of the drone be wherein denotes the current three-dimensional spatial position vector, denotes the current velocity vector, denotes the current attitude angle vector; Let the target state at the corresponding time in the optimal control path be wherein denotes a three-dimensional spatial position vector of the optimal control path, denotes a velocity vector of the optimal control path, denotes an attitude angle vector of the optimal control path; Calculate the state error, including: Position error represents the Euclidean distance between the current position and the target position; speed error , representing the modulus difference between the current speed and the target speed; pose error , representing an angular difference between the target pose and the current pose; and obtain disturbance intensity of current flight environment ; a preset position error tolerance a preset speed error tolerance a preset attitude error tolerance a disturbance intensity upper threshold and generates an anti-interference control instruction according to a set control instruction discrimination rule. 9.The urban logistics oriented UAV adaptive anti-interference precision landing control system of claim 8, wherein, The setting control instruction discrimination rule is specifically: If the following conditions are met , , and , the base control instruction is output, and the attitude and thrust are adjusted according to the optimal control path; If any of the state errors exceeds the corresponding tolerance or the disturbance intensity then the anti-disturbance compensation control mechanism is triggered, dynamically adjusting the control parameters and adding a feedforward disturbance compensation term; If the control parameter exceeds the upper limit for a plurality of control cycles in succession , and , and is above for a long time, the aircraft enters a protection mode in which the response sensitivity is limited to prevent loss of stability in flight. The finally generated anti-interference control instructions include attitude angle adjustment amount, thrust vector distribution value and time synchronization control parameter, and are sent to the flight control system to drive the UAV to execute anti-interference flight according to the optimal control path. 10.The urban logistics oriented UAV adaptive anti-interference precision landing control system of claim 2, wherein, According to the feedback information of the UAV after executing the anti-interference control instructions, the state evolution distribution of the fractal intention tensor body is updated in real time, specifically: In each control period, the actual flight feedback information of the UAV is collected, including the current position, speed, attitude angle and acceleration state data, and the feedback information is compared with the target state in the optimal control path in time sequence to obtain the deviation change sequence between the actual state and the target state; According to the deviation change sequence, the state error evolution trend is constructed, and combined with the environmental disturbance modeling information of the current region, the error evolution trend is fused and coded with the disturbance mode as a dynamic correction signal input to the asymmetric self-evolution intention generation network; In the fractal intention tensor body, the time slice tensor corresponding to the current control period is located and replaced with the dynamic correction signal, so as to complete the local intention update at the current time.

Images