Unmanned aerial vehicle autonomous navigation path planning system for slope inspection based on fusion neural network

Abstract

The application relates to the technical field of unmanned aerial vehicle path planning, and discloses an unmanned aerial vehicle side slope inspection autonomous navigation path planning system fusing a neural network, which comprises the following: a perception module acquires a continuous optical pixel offset matrix of an unmanned aerial vehicle and a three-axis disturbance state vector; an mapping module converts the optical pixel offset matrix into hidden space features to calculate error variance parameters, and generates a dynamic feedback damping gain factor; a control module determines adaptive weights according to the three-axis disturbance state vector to reset a track updating frequency and a repulsive field step length, and injects the dynamic feedback damping gain factor into a cost function to recalculate a value of the cost function, and generates an autonomous navigation instruction; the application fuses air flow disturbance parameters into a bottom layer track search logic, eliminates time domain misplacement of data correction and control decision, realizes feedforward compensation on high-frequency motion distortion, and suppresses navigation instruction lag divergence.

Description

Technical Field

[0001] This invention belongs to the field of UAV path planning technology, and in particular relates to an autonomous navigation path planning system for UAV slope inspection that integrates neural networks. Background Technology

[0002] Currently, the most common autonomous navigation method relies on airborne visual sensors to acquire spatial coordinates and then uses the controller to generate inspection tracks. However, when UAVs operate close to complex slope surfaces, canyon shear winds and updrafts on the slope will exert high-frequency random disturbances on the aircraft, causing motion distortion in the collected displacement data. To maintain track accuracy, conventional technologies usually adopt an asynchronous decoupled operating architecture between perception layer data correction and control layer path decision-making. The front-end computing module filters out image noise and then sends the smoothed coordinate data to the path planning operator.

[0003] This improvement path, which relies on enhancing the front-end filtering computing power to achieve accurate coordinate restoration, deviates from the inherent physical time lag of the flight control system. When the sensing link outputs corrective coordinates, the aircraft attitude has already undergone a secondary shift due to continuous airflow impact. For example, Chinese invention patent CN120353242B discloses a method for inspecting high slopes using a UAV and the UAV itself. It generates an initial flight path by reading design drawings and performs static or quasi-static corrections to waypoint positions based on slope flatness and laser ranging results. Analysis shows that the existing technology mainly suffers from… The system has the following shortcomings: 1. There is a spatiotemporal misalignment between the perception feedback flow and the action response chain, causing the path planner to continuously receive outdated environmental truth values, and the system loses its predictive compensation capability when faced with sudden airflow disturbances; 2. The controller forcibly tracks the filtered and noise-reduced hysteresis coordinates, which arouses navigation command oscillations under the action of shear wind, inducing dangerous drift of the inspection track towards the slope surface; 3. The system lacks an interactive loop that converts airflow disturbance energy into control feedback damping, and can only respond to high-frequency disturbances with a fixed search step size, making it difficult to autonomously balance the trajectory accuracy of near-field inspection with the safety of fuselage collision avoidance.

[0004] Therefore, the technical problem to be solved by this invention is how to break the operational architecture of the separation between perception filtering and control decision-making, directly translate the environmental distortion residual into the adaptive damping gain of the path planning operator, and block the track instability chain caused by command hysteresis from the physical feedback level. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A drone slope inspection autonomous navigation path planning system integrating neural networks, comprising:

[0006] First spatial physical quantity sensing module;

[0007] The second disturbance parameter mapping module is connected to the first spatial physical quantity sensing module;

[0008] And the third trajectory decision control module, which is connected to the second disturbance parameter mapping module and the first spatial physical quantity sensing module;

[0009] The first spatial physical quantity sensing module is configured to acquire the continuous optical pixel offset matrix and the three-axis disturbance state vector of the UAV in the physical space.

[0010] The second perturbation parameter mapping module is configured to receive the continuous optical pixel offset matrix, convert the continuous optical pixel offset matrix into latent space tensor features using a preset dimension reduction mapping network, calculate the dispersion of the latent space tensor features along the time dimension to output the error variance parameter, and generate a dynamic feedback damping gain factor based on the error variance parameter.

[0011] The third trajectory decision control module is configured to receive a dynamic feedback damping gain factor and a three-axis disturbance state vector, extract the frequency domain energy integral value of the three-axis disturbance state vector as a representative value of the motion energy level distribution, divide the representative value of the motion energy level distribution by a preset limit tolerance energy level value to output adaptive weights; multiply the preset base update frequency of the local path planning algorithm by the adaptive weights to obtain the reset track update frequency, and multiply the base step size of the local path planning algorithm by the adaptive weights to obtain the reset repulsion field step size; multiply the dynamic feedback damping gain factor by the current generation value of the local path planning algorithm to output the recalculated generation value, and generate autonomous navigation flight commands for the UAV based on the reset track update frequency, the reset repulsion field step size, and the recalculated generation value.

[0012] Preferably, the second perturbation parameter mapping module includes a time series construction submodule and a variance estimation submodule. The time series construction submodule is configured to receive a continuous optical pixel offset matrix and concatenate the continuous optical pixel offset matrix into a fixed-length sliding time window tensor. The variance estimation submodule is configured to receive the sliding time window tensor, input the sliding time window tensor into a long short-term memory network structure to generate a temporal feature tensor for the current time period, calculate the discrete distribution matrix of the temporal feature tensor, and extract the values ​​of the main diagonal elements of the discrete distribution matrix as the error variance parameter.

[0013] Preferably, the step of the second disturbance parameter mapping module generating a dynamic feedback damping gain factor based on the error variance parameter includes: processing the error variance parameter based on a preset exponential decay mathematical relationship; the exponential decay mathematical relationship is: Where D is the dynamic feedback damping gain factor, V is the error variance parameter, and k is the preset environmental response baseline coefficient.

[0014] Preferably, the steps of the third trajectory decision control module to extract the frequency domain energy integral value of the three-axis disturbance state vector as the representative value of the motion energy level distribution include: step 401, extracting the high-frequency vibration component of the three-axis disturbance state vector in the three-dimensional spatial coordinate system; step 402, performing a fast Fourier transform on the high-frequency vibration component to generate a frequency domain energy spectrum; step 403, calculating the integral area value of the frequency domain energy spectrum in the frequency band higher than the preset vibration frequency threshold, and setting the integral area value as the frequency domain energy integral value.

[0015] Preferably, in response to the adaptive weight being greater than a preset safety threshold parameter, the third trajectory decision control module executes frequency suppression logic; the frequency suppression logic includes: setting the reset track update frequency to the product of the base update frequency and 0.5, and setting the reset repulsion field step size to the product of the base step size and 0.33.

[0016] Preferably, the third trajectory decision control module is internally configured with a disturbance evolution tracking submodule; the disturbance evolution tracking submodule records the time series set of adaptive weights over multiple consecutive sampling periods and calculates the rate of change of the time derivative of the time series set within a fixed time sliding window; in response to the rate of change of the time derivative being greater than 0 for three consecutive sampling periods, the third trajectory decision control module outputs a hovering blocking signal and generates a hovering and waiting command that keeps the UAV in place with its current three-dimensional spatial coordinates, in lieu of the autonomous navigation flight command.

[0017] Preferably, the first spatial physical quantity sensing module includes an optical feature capture component and a spatial inertial measurement component; the optical feature capture component and the spatial inertial measurement component are connected through a hardware crystal oscillator clock synchronization trigger module; when the optical feature capture component receives a synchronization trigger signal, it locks the exposure to generate a digital image frame containing a continuous optical pixel offset matrix; when the spatial inertial measurement component receives a synchronization trigger signal, it synchronously records the triaxial acceleration values ​​and triaxial angular velocity values, and concatenates the triaxial acceleration values ​​and triaxial angular velocity values ​​into a triaxial disturbance state vector.

[0018] Preferably, the local path planning algorithm is an artificial potential field algorithm, and the current generation value is calculated by the gravitational potential field function and the repulsive potential field function. When the third trajectory decision control module outputs the recalculated generation value, it multiplies the dynamic feedback damping gain factor only with the output value of the gravitational potential field function, and keeps the output weight value of the repulsive potential field function constant.

[0019] Preferably, the third trajectory decision control module further includes a spatial limit boundary protection submodule; the spatial limit boundary protection submodule acquires the three-dimensional straight-line distance value between the UAV and the physical surface of the slope; in response to the three-dimensional straight-line distance value being less than 2m, the spatial limit boundary protection submodule outputs a coverage command, and sets the output weight value of the repulsive potential field function to 5 times the original value according to the coverage command.

[0020] Preferably, the environmental response baseline coefficient is set to 0.05 to 0.15; the digital image frame acquisition period is set to 20ms to 50ms; and the upper limit of the output frequency of autonomous navigation flight commands is set to 10Hz.

[0021] Compared to existing technologies, the UAV slope inspection autonomous navigation path planning system integrating neural networks has the following advantages: In the autonomous navigation path planning of UAV slope inspection, the physical coordination mechanism of the perception link and the control link is reconstructed. The system transforms the disturbance features representing continuous image pixel shifts into dynamic feedback damping gain factors within the control logic and injects them into the cost function update stage of the path planning algorithm. This architecture design deeply integrates environmental airflow disturbance parameters with the underlying trajectory search logic, eliminating the temporal misalignment between data correction and control decisions in the traditional asynchronous decoupling mode. The path planning module obtains information about complex airflow through this gain mechanism. The system's ability to predict impacts enables feedforward compensation for high-frequency motion distortion at the control feedback level. It also suppresses the hysteresis divergence of navigation commands by dynamically reconstructing the path search operator. The processor extracts the motion energy level distribution of the body disturbance state vector to generate adaptive weights and maps them to the path search step size operator in real time. When faced with increased environmental disturbance intensity, the system actively reduces the track update frequency and shrinks the repulsive field step size based on the energy level distribution. This state adjustment closed loop prevents the navigation controller from mechanically tracking discrete coordinate fluctuations containing motion distortion. Instead, it adds an anti-disturbance damping component to the track smoothness, blocking track drift induced by outdated truth values ​​from the physical action level. Attached Figure Description

[0022] Figure 1 This is a flowchart of the autonomous navigation command generation logic based on feature mapping and adaptive damping of the present invention.

[0023] Figure 2 This is a logical architecture diagram of the UAV autonomous navigation system integrating feedforward damping compensation according to the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0025] An autonomous navigation path planning system for unmanned aerial vehicle (UAV) slope inspection integrating neural networks includes:

[0026] First spatial physical quantity sensing module;

[0027] The second disturbance parameter mapping module is connected to the first spatial physical quantity sensing module;

[0028] And the third trajectory decision control module, which is connected to the second disturbance parameter mapping module and the first spatial physical quantity sensing module;

[0029] The first spatial physical quantity sensing module is configured to acquire the continuous optical pixel offset matrix and the three-axis disturbance state vector of the UAV in the physical space.

[0030] The second perturbation parameter mapping module is configured to receive the continuous optical pixel offset matrix, convert the continuous optical pixel offset matrix into latent space tensor features using a preset dimension reduction mapping network, calculate the dispersion of the latent space tensor features along the time dimension to output the error variance parameter, and generate a dynamic feedback damping gain factor based on the error variance parameter.

[0031] The third trajectory decision control module is configured to receive a dynamic feedback damping gain factor and a three-axis disturbance state vector, extract the frequency domain energy integral value of the three-axis disturbance state vector as a representative value of the motion energy level distribution, divide the representative value of the motion energy level distribution by a preset limit tolerance energy level value to output adaptive weights; multiply the preset base update frequency of the local path planning algorithm by the adaptive weights to obtain the reset track update frequency, and multiply the base step size of the local path planning algorithm by the adaptive weights to obtain the reset repulsion field step size; multiply the dynamic feedback damping gain factor by the current generation value of the local path planning algorithm to output the recalculated generation value, and generate autonomous navigation flight commands for the UAV based on the reset track update frequency, the reset repulsion field step size, and the recalculated generation value.

[0032] Preferably, the second perturbation parameter mapping module includes a time series construction submodule and a variance estimation submodule. The time series construction submodule is configured to receive a continuous optical pixel offset matrix and concatenate the continuous optical pixel offset matrix into a fixed-length sliding time window tensor. The variance estimation submodule is configured to receive the sliding time window tensor, input the sliding time window tensor into a long short-term memory network structure to generate a temporal feature tensor for the current time period, calculate the discrete distribution matrix of the temporal feature tensor, and extract the values ​​of the main diagonal elements of the discrete distribution matrix as the error variance parameter.

[0033] Preferably, the step of the second disturbance parameter mapping module generating a dynamic feedback damping gain factor based on the error variance parameter includes: processing the error variance parameter based on a preset exponential decay mathematical relationship; the exponential decay mathematical relationship is: Where D is the dynamic feedback damping gain factor, V is the error variance parameter, and k is the preset environmental response baseline coefficient.

[0034] Preferably, the steps of the third trajectory decision control module to extract the frequency domain energy integral value of the three-axis disturbance state vector as the representative value of the motion energy level distribution include: step 401, extracting the high-frequency vibration component of the three-axis disturbance state vector in the three-dimensional spatial coordinate system; step 402, performing a fast Fourier transform on the high-frequency vibration component to generate a frequency domain energy spectrum; step 403, calculating the integral area value of the frequency domain energy spectrum in the frequency band higher than the preset vibration frequency threshold, and setting the integral area value as the frequency domain energy integral value.

[0035] Preferably, in response to the adaptive weight being greater than a preset safety threshold parameter, the third trajectory decision control module executes frequency suppression logic; the frequency suppression logic includes: setting the reset track update frequency to the product of the base update frequency and 0.5, and setting the reset repulsion field step size to the product of the base step size and 0.33.

[0036] Preferably, the third trajectory decision control module is internally configured with a disturbance evolution tracking submodule; the disturbance evolution tracking submodule records the time series set of adaptive weights over multiple consecutive sampling periods and calculates the rate of change of the time derivative of the time series set within a fixed time sliding window; in response to the rate of change of the time derivative being greater than 0 for three consecutive sampling periods, the third trajectory decision control module outputs a hovering blocking signal and generates a hovering and waiting command that keeps the UAV in place with its current three-dimensional spatial coordinates, in lieu of the autonomous navigation flight command.

[0037] Preferably, the first spatial physical quantity sensing module includes an optical feature capture component and a spatial inertial measurement component; the optical feature capture component and the spatial inertial measurement component are connected through a hardware crystal oscillator clock synchronization trigger module; when the optical feature capture component receives a synchronization trigger signal, it locks the exposure to generate a digital image frame containing a continuous optical pixel offset matrix; when the spatial inertial measurement component receives a synchronization trigger signal, it synchronously records the triaxial acceleration values ​​and triaxial angular velocity values, and concatenates the triaxial acceleration values ​​and triaxial angular velocity values ​​into a triaxial disturbance state vector.

[0038] Preferably, the local path planning algorithm is an artificial potential field algorithm, and the current generation value is calculated by the gravitational potential field function and the repulsive potential field function. When the third trajectory decision control module outputs the recalculated generation value, it multiplies the dynamic feedback damping gain factor only with the output value of the gravitational potential field function, and keeps the output weight value of the repulsive potential field function constant.

[0039] Preferably, the third trajectory decision control module further includes a spatial limit boundary protection submodule; the spatial limit boundary protection submodule acquires the three-dimensional straight-line distance value between the UAV and the physical surface of the slope; in response to the three-dimensional straight-line distance value being less than 2m, the spatial limit boundary protection submodule outputs a coverage command, and sets the output weight value of the repulsive potential field function to 5 times the original value according to the coverage command.

[0040] Preferably, the environmental response baseline coefficient is set to 0.05 to 0.15; the digital image frame acquisition period is set to 20ms to 50ms; and the upper limit of the output frequency of autonomous navigation flight commands is set to 10Hz.

[0041] Example 1: When a UAV performs close-range inspection of a canyon slope with an elevation difference greater than 300m, the canyon shear wind and the updraft on the slope alternately act on the fuselage. After locking the exposure, the optical feature capture component in the airborne first spatial physical quantity perception module generates a digital image frame. This digital image frame is mixed with nonlinear motion distortion caused by high-frequency aerodynamic disturbances. Under the conventional architecture of asynchronous decoupling between the perception link and the control link, the action of the front-end computing module to filter out image noise will inevitably cause a spatiotemporal misalignment between the perception feedback flow and the action response chain. This causes the third trajectory decision control module to continuously receive the environmental truth with physical time lag, thereby arousing navigation command oscillation and inducing the UAV to shift towards the slope surface.

[0042] To address the aforementioned airflow disturbance conditions, this technical solution reconstructs the underlying data correction logic into a feedforward control gain mechanism that directly acts on the path planning operator. Specifically, in the operational process, the first spatial physical quantity sensing module acquires the continuous optical pixel offset matrix and the three-axis disturbance state vector, and synchronously transmits them to the second disturbance parameter mapping module and the third trajectory decision control module. The second disturbance parameter mapping module uses a preset dimensionality reduction mapping network to convert the continuous optical pixel offset matrix into latent space tensor features, calculates the dispersion of the latent space tensor features along the time dimension to output the error variance parameter V, and the second disturbance parameter mapping module is based on the exponential decay mathematical relationship... The dynamic feedback damping gain factor D is generated, where k is a preset environmental response benchmark coefficient. The third trajectory decision control module extracts the high-frequency vibration components of the three-axis disturbance state vector in the three-dimensional spatial coordinate system, performs a fast Fourier transform on the high-frequency vibration components to generate a frequency domain energy spectrum, calculates the integral area of ​​the frequency domain energy spectrum in the frequency band above the preset vibration frequency threshold as the frequency domain energy integral value, and extracts it as the representative value of the motion energy level distribution. The third trajectory decision control module divides the representative value of the motion energy level distribution by the preset limit bearing energy level value and outputs the adaptive weight.

[0043] As the intensity of airflow disturbance increases, the error variance parameter V increases, and the dynamic feedback damping gain factor D calculated according to the above exponential decay mathematical relationship decreases accordingly. The third trajectory decision control module multiplies the dynamic feedback damping gain factor D only with the output value of the gravitational potential field function, while keeping the output weight value of the repulsive potential field function constant, and outputs the recalculated cost value. At the same time, the third trajectory decision control module multiplies the basic update frequency of the local path planning algorithm with the adaptive weight to obtain the reset track update frequency, and multiplies the basic step size of the local path planning algorithm with the adaptive weight to obtain the reset repulsive field step size. The third trajectory decision control module generates the UAV's autonomous navigation flight command based on the reset track update frequency, the reset repulsive field step size, and the recalculated cost value. The system relies on the cross-module state transition mechanism to translate the environmental distortion residual into the adaptive damping gain of the path planning operator, actively reducing the gravitational field tracking frequency of the target waypoint and shrinking the repulsive field detection step size, thereby blocking the track drift chain induced by outdated truth values ​​from the physical action level and maintaining the global stability of the near-field inspection track.

[0044] Example 2: A hardware-in-the-loop simulation test platform was constructed to test the control stability of the dynamic feedback damping gain mechanism under the condition of UAV trajectory drift induced by canyon shear wind. A six-degree-of-freedom UAV flight dynamics model and a computational fluid dynamics wind field model were integrated. The wind field model solved the Navier-Stokes equations and output a continuous three-dimensional airflow disturbance field. To reproduce the mechanical vibration of the optical sensor in physical space, Gaussian white noise with a signal-to-noise ratio of 15dB was injected into the input link of the optical pixel offset matrix, and periodic power frequency interference with a frequency of 50Hz was superimposed. The base update frequency of the local path planning algorithm was set according to the environmental disturbance spectrum bandwidth and the computational load of the airborne processor. Fixed high-frequency path updates caused instruction accumulation and actuator overload. The system set the base update frequency benchmark value to 50Hz. The third trajectory decision control module used adaptive weights to multiply the base update frequency so that the trajectory update frequency decreased when the disturbance energy crossed the preset vibration frequency threshold.

[0045] The simulation test platform was launched and the shear wind speed was set to 12.0 m / s. Simultaneously, the root mean square error (RMSE) of the UAV's three-dimensional coordinate tracking and the minimum distance from the slope surface were measured. The first control group had the dynamic feedback damping gain factor D removed, while the adaptive weights were retained. The second control group had the adaptive weights removed, while the dynamic feedback damping gain factor D was retained. The experimental group was equipped with a complete dimensionality reduction mapping network and parameter adjustment mechanism. The environmental response baseline coefficient k was set to 1.0. At the moment of strong wind shear access, the first control group showed an RMS error that increased to 2.45 m, and the minimum distance from the slope surface decreased to 0.85 m. The second control group showed a RMS error... The difference was 1.98m, the navigation command response delay was measured to be 145.2ms, the root mean square error of the test group converged to 0.62m, the minimum distance from the slope surface was maintained at 2.34m, and the command response delay was reduced to 42.1ms. The original flight trajectory of the unprocessed first control group showed high-frequency irregular sawtooth oscillations, while the trajectory of the test group showed a smooth damped following curve. The test data showed that the dynamic feedback damping gain factor D suppressed the gravitational potential field tracking tendency, and the adaptive weight synchronously contracted the repulsive field detection step size. The two parameters constructed a feedforward suppression closed loop at the physical action level, blocking the command oscillation induced by the environmental truth hysteresis.

[0046] To establish the working boundary of the environmental response reference coefficient k, a parameter gradient control group was configured. The environmental response reference coefficient k was set to 0.05, 1.0, and 5.5, respectively. The extreme values ​​of track deviation for each group were measured in a continuous airflow impact sequence. When the environmental response reference coefficient k was set to the lower limit of 0.05, the error variance parameter V was insufficient to drive the exponential decay mathematical relationship to generate sufficient damping, and the measured extreme value of track deviation reached 1.85m. When the environmental response reference coefficient k was set to 1.0, the extreme value of track deviation decreased and remained at 0.58m. When the environmental response reference coefficient k was set to the upper limit of 5.5, the dynamic feedback damping gain factor D output by the exponential decay mathematical relationship showed a nonlinear accelerated decrease, and the system entered an overdamped stop. In the stagnant state, the UAV loses its ability to track and maneuver the target waypoint, and the extreme value of the track deviation climbs to 3.75m. The data curve shows a deterioration inflection point. Test data shows that the environmental response baseline coefficient k is in the range of 0.8 to 1.2, which achieves a technical balance between anti-disturbance stability and tracking sensitivity. The continuous optical pixel offset matrix is ​​translated into a dynamic feedback damping gain factor D through a dimension reduction mapping network. In conjunction with the adaptive weight calculated based on the three-axis disturbance state vector, the numerical distribution of the underlying path cost function is changed. The nonlinear motion distortion caused by high-frequency aerodynamic disturbance is directly converted into feedforward damping parameters, which solves the technical problem between coordinate tracking accuracy and global flight collision avoidance safety, and maintains the global stability of the UAV's inspection track in wind field environment.

[0047] Example 3: This example combines Figures 1 to 2 This section describes an autonomous navigation path planning system for UAV slope inspection that incorporates neural networks. Figure 1 As shown, the first spatial physical quantity sensing module acquires physical space UAV data and outputs a continuous optical pixel offset matrix and a three-axis perturbation state vector downstream. The continuous optical pixel offset matrix enters the second perturbation parameter mapping module, which processes the features using a preset dimensionality reduction mapping network to sequentially generate latent space tensor features (obtained through the preset dimensionality reduction mapping network transformation). The error variance parameter is extracted, which is obtained by calculating the dispersion along the time dimension. Finally, the dynamic feedback damping gain factor is output, which is generated based on the error variance parameter. Simultaneously, the three-axis perturbation state vector enters the third trajectory decision control module, which makes command decisions based on the perturbation parameters and obtains the motion energy level distribution. The representative value is obtained by extracting the frequency domain energy integral value, and then the adaptive weight is obtained by dividing it by the preset limit energy level value. This adaptive weight is used to calculate the reset repulsion field step size (the algorithm's basic step size multiplied by the adaptive weight) and the reset track update frequency (the algorithm's basic update frequency multiplied by the adaptive weight). In addition, the aforementioned generated dynamic feedback damping gain factor is imported to obtain the recalculated cost value, which is multiplied by the current cost value of the local path planning algorithm. Finally, the parameters of the above branches are gathered to generate autonomous navigation flight commands, which are generated based on the update frequency, repulsion field step size, and recalculated cost value. Finally, the circular primitive node that executes the action and provides feedback is pointed to by an arrow.

[0048] like Figure 2 As shown, the state monitoring results generated by the continuous operation of the system are objectively presented in a two-dimensional distribution system. The horizontal axis represents time and is calibrated in seconds, with a value range of 0 to 60. The vertical axis represents displacement and is calibrated in meters, with a value range of -3.0 to 3.0. The legend clearly distinguishes the two sets of test data sequences, including the first control group and the experimental group. In this coordinate system, the displacement curve of the first control group shows an oscillating pattern that alternates significantly between the positive and negative ranges of the vertical axis when extended on the time axis. In contrast, the displacement curve of the experimental group closely follows the zero point region of the horizontal axis, showing a convergent state with reduced fluctuation amplitude and an overall tendency to flatten.

[0049] Example 4: The UAV is in a slope inspection wind field environment with nonlinear motion distortion. The dimensionality reduction mapping network in the second perturbation parameter mapping module includes a spatial convolutional layer and a long short-term memory network structure. The second perturbation parameter mapping module inputs the acquired continuous optical pixel offset matrix into the spatial convolutional layer. The spatial convolutional layer consists of 3 layers, each configured with 16 two-dimensional convolution operators of size 3x3 and stride 1. Each convolutional layer is followed by a max pooling layer with stride 2, compressing the original image into a 16x16 spatial dimension feature map. The network structure contains 64 hidden units and is connected to a long short-term memory network structure. Sequence correlation is extracted through time step iteration, and the spatial dimension feature map is calculated and output. The spatial dimension feature maps of a preset number of sampling periods are spliced ​​along the time axis. The specific sampling strategy is as follows: a first-in-first-out circular buffer with a depth of 30 sampling periods is set. When each frame of feature map is enqueued, the buffer is discarded. The oldest frame of data is used, and the overlap rate of two adjacent sliding windows in the buffer is kept at 50% to ensure sufficient temporal smoothness within the acquisition period of 20ms to 50ms, thereby constructing a sliding time window tensor. The second perturbation parameter mapping module inputs the sliding time window tensor into the long short-term memory network structure to generate the temporal feature tensor of the current time period. The element-wise subtraction of two adjacent feature vectors in the temporal feature tensor is performed and the absolute value is taken. The resulting absolute values ​​are accumulated into a 25x25 square matrix, so that the main diagonal elements of the square matrix record the feature fluctuation of adjacent time steps. The 25 accumulated values ​​on the main diagonal are extracted, and their arithmetic mean is calculated and defined as the error variance parameter to reflect the time delay uncertainty of image features within the sliding window. At the same time, the discrete distribution matrix of the temporal feature tensor is calculated, and the values ​​of the main diagonal elements of the discrete distribution matrix are extracted as the error variance parameter.

[0050] For the limit energy level value invoked by the third trajectory decision control module, an offline calibration procedure is executed based on the hardware-in-the-loop platform. The UAV carrying the standard inspection payload is fixed on the wind tunnel test platform. The wind tunnel generator outputs a shear gust sequence with increasing flow velocity. The airborne sensing component synchronously acquires the UAV's three-axis disturbance state vector. The third trajectory decision control module extracts the high-frequency vibration components of the three-axis disturbance state vector in the three-dimensional spatial coordinate system and performs a fast Fourier transform. The integral area of ​​the frequency domain energy spectrum in the frequency band above the preset vibration frequency threshold is calculated. The numerical output is the real-time frequency domain energy integral value. The calibration steps for the ultimate withstand energy level value are as follows: In the initial preparation stage of the UAV power-on, the wind shear speed is increased in step increments of 0.5 m / s in the control wind tunnel or simulated environment. The yaw angle fluctuation of the fuselage is monitored in real time. When the single yaw angle deviation reaches 12.0 degrees for the first time or the PWM duty cycle of the drive motor reaches the physical saturation threshold of 85%, the real-time frequency domain energy integral value calculated by the current processor is taken as the calibration benchmark. The test is repeated for 8 sets, and the arithmetic mean is stored in the underlying register. The limit energy level value is set as the limit energy level value. The real-time frequency domain energy integral value is divided by the limit energy level value in the register to obtain a dimensionless ratio between 0.2 and 1.0, which is then directly assigned to the adaptive weight for subsequent linear scaling control of the trajectory update frequency. During the process of increasing gust speed, when the UAV's fuselage yaw angle exceeds the preset safety angle threshold once, or when the output power of the underlying drive motor reaches the rated full load state, the system locks the current timestamp, extracts the real-time frequency domain energy integral value corresponding to the timestamp, and sets it as the limit energy level value. Based on the tensor calculation process of the continuous optical pixel offset matrix to the error variance parameter and the physical calibration procedure of the limit energy level value based on the wind tunnel test platform, the third trajectory decision control module generates a dynamic feedback damping gain factor and an adaptive weight according to the extracted error variance parameter and the calibrated limit energy level value. The UAV receives the reset trajectory update frequency and reset repulsive field step size determined by the adaptive weight, and controls the physical actuator to move according to the autonomous navigation command containing the above parameters in the slope inspection condition.

[0051] Example 5: When the preset vibration frequency threshold is determined, the offline calibration procedure is initiated based on the inherent mechanical resonance characteristics of the fuselage. The UAV is rigidly connected to the variable frequency vibration test bench, and the test bench is controlled to output a continuous logarithmic sweep frequency excitation signal. The airborne first spatial physical quantity sensing module synchronously collects the three-axis vibration response sequence of the UAV under the sweep frequency excitation. The front-end calculation module calculates the amplitude-frequency characteristic curve of the three-axis vibration response sequence, extracts the initial frequency node where the amplitude response crosses the steady-state baseline and exhibits nonlinear amplification as the lower boundary of structural resonance, and the system control unit writes the frequency value corresponding to the lower boundary of structural resonance into the bottom register and locks it as the preset vibration frequency threshold. The third trajectory decision control module separates the inherent mechanical vibration of the fuselage from the external wind field disturbance according to the preset vibration frequency threshold and outputs the frequency domain energy integral value corresponding to the external airflow impact energy level.

[0052] When the UAV is in the takeoff preparation state of starting the rotor blades, the system initiates the on-site pre-calibration procedure for the environmental response reference coefficient k based on the current physical load status. The UAV maintains an idling state on the ground and locks the basic position coordinates. The second disturbance parameter mapping module collects the initial optical pixel offset matrix under static windless conditions and generates the baseline error variance parameter using a dimension reduction mapping network. The system control unit calculates the logarithmic ratio of the preset target initial damping gain boundary value to the baseline error variance parameter to output the reference coefficient. The system control unit assigns this reference coefficient as the environmental response reference coefficient k. Based on the assigned environmental response reference coefficient k, the system control unit filters out the background static noise introduced by the rotor blade idling vibration, locks the trigger zero point of the dynamic feedback damping gain factor, and maintains the control stability of the feedforward damping parameter in the initial stage of slope inspection.

[0053] To determine the parameters of the continuous optical pixel offset matrix processed by the second disturbance parameter mapping module, the dimensionality reduction mapping network, in its offline state, uses a windless environment to collect a calibration displacement sequence and a preset amplitude disturbed displacement sequence to complete supervised learning. The spatial feature extraction layer converts the optical pixel offset matrix into an N-dimensional feature vector. The long short-term memory network structure extracts the temporal correlation of the feature vector within a preset length sliding time window. The generated temporal feature tensor is used to calculate the error variance parameter V through the main diagonal elements of the discrete distribution matrix. To ensure that the output of the dynamic feedback damping gain factor D meets the preset control accuracy requirements, a static noise benchmark is obtained through an on-site pre-calibration procedure. The initial optical pixel offset matrix is ​​collected at the propeller idle speed before the UAV takes off, and the corresponding baseline error variance parameter is calculated. The environmental response benchmark coefficient k is determined based on the logarithmic ratio of the target initial damping gain boundary value to this parameter, locking the trigger zero point of the dynamic feedback damping gain factor and eliminating the measurement deviation introduced by the inherent vibration of the power system.

[0054] Example 6: When a UAV is deployed in a slope environment expected to include extreme gust wind shear, the increase in adaptive weights output by the third trajectory decision control module leads to a saturation risk in the underlying actuators due to high-frequency command scheduling. To address this, the system initiates an offline boundary calibration procedure for safety threshold parameters before formally executing the inspection task. Technicians place the UAV in a wind tunnel chamber with a controllable wind speed and continuously increase the wind dynamic pressure output by the wind tunnel at a set step-by-step incremental rate. The airborne first spatial physical quantity sensing module and the third trajectory decision control module calculate and output the corresponding adaptive weights in real time based on the aforementioned hardware-in-the-loop mapping logic. According to the weight sequence, the monitoring system simultaneously collects the three-phase transient current and PWM duty cycle parameters of each rotor drive motor of the UAV. During the process of increasing wind pressure, when the PWM duty cycle of any single-axis drive motor reaches the physical saturation critical point of 95% for 5 consecutive sampling cycles, the test system extracts the corresponding adaptive weight value output by the third trajectory decision control module at that time stamp. The system multiplies the extracted adaptive weight value by an engineering anti-shake redundancy coefficient of 0.8 to obtain the product value and directly sets it as the preset safety threshold parameter, thus completing the solidification and writing of the safety threshold parameter in the underlying control register.

[0055] In real-world slope inspection route control, the third trajectory decision control module extracts the three-axis disturbance state vector in real time and dynamically calculates adaptive weights. When the real-time value of the adaptive weights exceeds the preset safety threshold parameter, the third trajectory decision control module triggers the hard constraint intervention of the frequency suppression logic. The third trajectory decision control module stops executing the conventional proportional attenuation calculation link, resets the trajectory update frequency to the product of the base update frequency and 0.5, and simultaneously increases and amplifies the reset repulsive field step size according to the preset multiple. The aforementioned safety threshold parameter obtained based on physical saturation characteristics, combined with the condition-triggered frequency reduction and multiplication operation, enables the UAV to enter a low-frequency conservative detection mode when facing extreme airflow impacts, blocking the cascade transmission path of high-frequency environmental noise to the attitude control link and maintaining the operational stability of the UAV's underlying power system in harsh environments.

[0056] When determining the ultimate energy level and safety threshold parameters of the third trajectory decision control module, hardware-in-the-loop simulation combined with physical wind tunnel testing is used to complete the quantitative calibration. Under the condition of the UAV carrying a standard inspection payload, the wind tunnel generator is controlled to output a flow rate increasing shear gust, and the high-frequency vibration components of the three-axis disturbance state vector are extracted in real time and a fast Fourier transform is performed. When the single deviation of the UAV fuselage yaw angle exceeds the preset safety angle threshold or the output power of any rotor drive motor reaches the rated full load state, the frequency domain energy integral value corresponding to the current timestamp is extracted and solidified as the ultimate energy level value. This value is multiplied by an engineering redundancy coefficient of 0.8 to determine the safety threshold parameter, which serves as the hard constraint boundary of the trigger frequency suppression logic. When the adaptive weight is calculated in real time and exceeds its boundary, the system forcibly reduces the reset track update frequency and shrinks the reset repulsive field step size, so that the UAV can maintain physical dynamic balance and inspection track smoothness by adjusting the path tracking sensitivity under extreme airflow conditions.

[0057] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A drone-based autonomous navigation path planning system for slope inspection integrating neural networks, characterized in that, include: First spatial physical quantity sensing module; The second disturbance parameter mapping module is connected to the first spatial physical quantity sensing module; And the third trajectory decision control module, which is connected to the second disturbance parameter mapping module and the first spatial physical quantity sensing module; The first spatial physical quantity sensing module is configured to acquire the continuous optical pixel offset matrix and the three-axis disturbance state vector of the UAV in the physical space. The second perturbation parameter mapping module is configured to receive the continuous optical pixel offset matrix, convert the continuous optical pixel offset matrix into latent space tensor features using a preset dimension reduction mapping network, calculate the dispersion of the latent space tensor features along the time dimension to output the error variance parameter, and generate a dynamic feedback damping gain factor based on the error variance parameter. The third trajectory decision control module is configured to receive the dynamic feedback damping gain factor and the three-axis disturbance state vector, extract the frequency domain energy integral value of the three-axis disturbance state vector as the representative value of the motion energy level distribution, and divide the representative value of the motion energy level distribution by the preset limit tolerance energy level value to output the adaptive weight. The preset local path planning algorithm's base update frequency is multiplied by the adaptive weights to obtain the reset track update frequency, and the local path planning algorithm's base step size is multiplied by the adaptive weights to obtain the reset repulsion field step size; the dynamic feedback damping gain factor is multiplied by the current generation value of the local path planning algorithm to output the recalculated generation value, and the UAV's autonomous navigation flight command is generated based on the reset track update frequency, the reset repulsion field step size, and the recalculated generation value. The second perturbation parameter mapping module includes a time series construction submodule and a variance estimation submodule. The time series construction submodule is configured to receive continuous optical pixel offset matrices and concatenate the continuous optical pixel offset matrices into a fixed-length sliding time window tensor. The variance estimation submodule is configured to receive a sliding time window tensor, input the sliding time window tensor into a long short-term memory network structure to generate a temporal feature tensor for the current time period, calculate the discrete distribution matrix of the temporal feature tensor, and extract the values ​​of the main diagonal elements of the discrete distribution matrix as the error variance parameter. The second perturbation parameter mapping module generates a dynamic feedback damping gain factor based on the error variance parameter, including: processing the error variance parameter based on a preset exponential decay mathematical formula; the exponential decay mathematical formula is: Where D is the dynamic feedback damping gain factor, V is the error variance parameter, and k is the preset environmental response benchmark coefficient; The first spatial physical quantity sensing module includes an optical feature capture component and a spatial inertial measurement component; the optical feature capture component and the spatial inertial measurement component are connected through a hardware crystal oscillator clock synchronization trigger module; when the optical feature capture component receives a synchronization trigger signal, it locks the exposure to generate a digital image frame containing a continuous optical pixel offset matrix; when the spatial inertial measurement component receives a synchronization trigger signal, it synchronously records the three-axis acceleration values ​​and the three-axis angular velocity values, and concatenates the three-axis acceleration values ​​and the three-axis angular velocity values ​​into a three-axis disturbance state vector.

2. The UAV slope inspection autonomous navigation path planning system integrating neural networks according to claim 1, characterized in that, The steps of the third trajectory decision control module to extract the frequency domain energy integral value of the three-axis disturbance state vector as the representative value of the motion energy level distribution include: step 401, extracting the high-frequency vibration component of the three-axis disturbance state vector in the three-dimensional spatial coordinate system; step 402, performing a fast Fourier transform on the high-frequency vibration component to generate a frequency domain energy spectrum; step 403, calculating the integral area value of the frequency domain energy spectrum in the frequency band higher than the preset vibration frequency threshold, and setting the integral area value as the frequency domain energy integral value.

3. The UAV slope inspection autonomous navigation path planning system integrating neural networks according to claim 1, characterized in that, In response to the adaptive weight value being greater than the preset safety threshold parameter, the third trajectory decision control module executes frequency suppression logic; the frequency suppression logic includes: setting the reset track update frequency to the product of the base update frequency and 0.5, and setting the reset repulsion field step size to the product of the base step size and 0.

33.

4. The UAV slope inspection autonomous navigation path planning system integrating neural networks according to claim 1, characterized in that, The third trajectory decision control module is internally configured with a disturbance evolution tracking submodule. The disturbance evolution tracking submodule records the time series set of adaptive weights over multiple consecutive sampling periods and calculates the rate of change of the time derivative of the time series set within a fixed time sliding window. In response to the rate of change of the time derivative being greater than 0 for three consecutive sampling periods, the third trajectory decision control module outputs a hovering blocking signal and generates a hovering and waiting command that keeps the UAV in place with its current three-dimensional spatial coordinates, replacing the autonomous navigation flight command.

5. The UAV slope inspection autonomous navigation path planning system integrating neural networks according to claim 1, characterized in that, The local path planning algorithm is an artificial potential field algorithm, and the current generation value is calculated by the gravitational potential field function and the repulsive potential field function. When the third trajectory decision control module outputs the recalculated generation value, it multiplies the dynamic feedback damping gain factor only with the output value of the gravitational potential field function, and keeps the output weight value of the repulsive potential field function constant.

6. The UAV slope inspection autonomous navigation path planning system integrating neural networks according to claim 5, characterized in that, The third trajectory decision control module also includes a spatial limit boundary protection submodule; the spatial limit boundary protection submodule obtains the three-dimensional straight-line distance value between the UAV and the physical surface of the slope; in response to the three-dimensional straight-line distance value being less than 2m, the spatial limit boundary protection submodule outputs a coverage command, and sets the output weight value of the repulsive potential field function to 5 times the original value according to the coverage command.

7. The UAV slope inspection autonomous navigation path planning system integrating neural networks according to claim 1, characterized in that, The environmental response baseline coefficient is set to 0.05 to 0.15; the digital image frame acquisition period is set to 20ms to 50ms; and the upper limit of the output frequency of autonomous navigation flight commands is set to 10Hz.

Images