A mine complex topography hyperspectral mapping stable posture control method and system

Abstract

The present application belongs to the technical field of remote sensing mapping, and discloses a kind of mine complex topography hyperspectral mapping stable posture control method and system.Collect flight platform body posture, angular velocity, gimbal posture, terrain elevation, ranging and vibration data, to complete multi-source data alignment with hyperspectral camera exposure time as unified time scale;Based on terrain elevation and ranging data, construct local surface elevation window, calculate terrain gradient and surface normal vector to determine ideal load visual axis reference, solve actual load visual axis to get visual axis error angle and scanning direction mismatch amount;Fusion vibration characteristics construct enhanced state vector and discrete prediction model, rolling solve gimbal compensation control sequence and superimpose fast feedback to generate instruction, switch security degradation control in abnormal time, support closed-loop parameter update and version rollback mechanism, effectively improve the stability of hyperspectral push-broom imaging in complex topography and strip mapping quality.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing and mapping technology, specifically relating to a method and system for attitude stabilization control of airborne hyperspectral push-broom mapping in complex terrain scenarios in mining areas. It can be applied to high-precision hyperspectral mapping operations in areas such as open-pit mines, tailings ponds, and slope steps. Background Technology

[0002] With the increasing demand for refined monitoring of open-pit mines, spoil heaps, slope steps, tailings ponds, and mining subsidence areas, airborne hyperspectral mapping is gradually becoming an important means of lithological identification, weathering zoning, water-bearing anomaly monitoring, and disaster hazard investigation in mining areas. Unlike ordinary visible light mapping, hyperspectral payloads typically employ a pushbroom imaging structure, where a single spectral line accumulates line by line to form a data cube as the platform advances. Therefore, its spatial and spectral dimensions are not acquired simultaneously, but rather rely on stable coupling along the time axis due to platform attitude, velocity, line-of-sight orientation, and exposure sequence. Once the UAV is subjected to aerodynamic interference from slope reflections, rotor downwash, canyon crosswinds, and abrupt aerodynamic changes at step edges over complex terrain, high-frequency perturbations in pitch, roll, and yaw will be mapped as slit direction drift, non-uniform ground projection velocity, and line-of-sight tangency errors. This leads to spectral band misalignment, pixel stretching, uneven overlap of adjacent scan lines, and even the mixing of spectra from different ground features. Such errors are not simply image jitter problems, but rather time-varying geometric mismatches between flight attitude, payload line-of-sight orientation, and terrain normal.

[0003] From a mechanistic perspective, the challenge in complex terrain scenarios within mining areas lies in the multi-timescale coupling characteristics of disturbance sources: high-frequency angular disturbances are caused by aircraft vibration and gimbal structural elasticity, mid-frequency attitude shifts are caused by terrain undulations and flight control loops, and changes in the field of view ground normal caused by abrupt slope changes manifest as low-frequency but spatially non-stationary observation geometric changes. Traditional fixed-attitude flight or methods relying solely on flight control attitude maintenance rely on stable aircraft attitude to meet the control prerequisite for stable mapping, without explicitly modeling the relationship between terrain slope changes, target normals on the line of sight, and gimbal compensation requirements. Methods relying solely on post-exposure geometric correction rely on the technical premise that attitude disturbances at the exposure time can be inversely compensated in post-processing. However, in pushbroom imaging, since each row of pixels corresponds to different instantaneous attitudes and different ground incidence angles, if line-of-sight stabilization is not controlled during the acquisition phase, post-processing can only correct some geometric distortions and cannot recover the cross-surface spectral mixing and modulation transfer function degradation that have already occurred.

[0004] To address the aforementioned issues, existing technologies have been explored to some extent. Based on the technical approach, they can be mainly divided into two categories: one focuses on self-compensation of the optical field of view within the imaging payload, and the other focuses on terrain perception and disturbance perception adjustment on the flight platform side.

[0005] The first category involves self-compensation of the optical field of view within the imaging payload. For example, Chinese patent CN202511892421.4 discloses a hyperspectral camera system and control method with self-stabilized and self-adjusted field of view. It acquires angular perturbations in real time and calculates the two-dimensional compensation displacement of the slit, driving the slit to move in the opposite direction to maintain a basically constant field of view. It also combines the mapping of the slit position to ideal coordinates to implement geometric transformation and dynamic correction through resampling. The control premise of this type of technology is that the main perturbations of the hyperspectral system can be equivalently compensated within the camera through slit translation. Its control object is the field of view position deviation on the image plane, and its action boundary is concentrated within the imaging link. The time scale is more biased towards the synchronization level of exposure and slit drive. It does not explicitly address the problem of reconstructing the optimal observation attitude caused by changes in the terrain normal of the mining area, nor does it establish a closed-loop cooperative relationship between the flight platform attitude, gimbal angle, and ground slope.

[0006] The second category involves terrain perception and disturbance perception adjustment on the flight platform side. For example, Chinese patent CN202610437604.5 discloses a collaborative monitoring adaptive cascade sprinkler control method, which uses terrain slope and aspect as rigid constraints to establish a three-dimensional mapping relationship between remote sensing image pixels and ground sensor measurement points, and uses a terrain-adaptive synchronous control law for local trajectory tracking. Chinese patent CN202610238199.4 discloses a rapid identification method for geological disaster hazard points based on UAV aerial surveying, which uses image exposure time as a unified time axis, integrates inertial measurement, propeller speed, wind speed and direction, and aircraft vibration signals, calculates the equivalent micro-attitude disturbance spectrum, and triggers an adaptive fine-tuning strategy for flight parameters. This type of technology is developed from the perspectives of terrain-constrained flight and disturbance identification, respectively. The control premise is that the adjustment of flight path or aircraft control parameters is sufficient to improve the mission effect, and the control focus is on the path scale or the exposure-related disturbance analysis scale, respectively. However, most of these technologies handle body attitude control and load line-of-sight control separately, without constructing dedicated state variables and compensation laws to address the requirement that the stability of the pushbroom hyperspectral line-of-sight with respect to the Earth's normal direction takes precedence over the stability of the body itself.

[0007] The aforementioned existing technologies have the following shortcomings in practical applications. First, the definition of the compensation object is flawed. The first type of method mainly treats the error as field-of-view displacement or slit position deviation within the image plane, with the control objective being to maintain the geometric stability of the camera's internal imaging. However, in hyperspectral mapping of complex terrain in mining areas, the key quantity truly affecting spectral availability is not simply image plane drift, but rather the time-varying deviation of the angle between the line-of-view direction and the local surface normal, as well as the mismatch between the scanning direction and the ground projection velocity. Compensating only the camera's internal field of view cannot suppress cross-line spectral mixing caused by aircraft roll, slope tilt, and abrupt changes in terrain normal. Second, the aircraft control and payload control architectures are separated. The second type of method emphasizes flight path and local tracking under terrain constraints, as well as fine-tuning of flight parameters triggered by disturbance evidence. Essentially, it still treats the aircraft as the primary control object, assuming the payload can passively follow the aircraft for stability. For hyperspectral pushbroom imaging, aircraft attitude stability is not equivalent to slit line-of-view stability, because arm vibration, gimbal mounting flexibility, payload centroid offset, and structural resonance can introduce phase differences between the aircraft and the payload. Existing methods lack a collaborative compensation mechanism that unifies aircraft attitude, terrain slope, vibration spectrum, and gimbal angle position into a single state space. Furthermore, they lack a predictive attitude stabilization mechanism oriented towards the pushbroom imaging time axis. Current technologies mostly perform reverse corrections or parameter fine-tuning after a disturbance occurs, with the control logic remaining at the post-observation compensation level. However, hyperspectral pushbroom imaging is extremely sensitive to exposure timing; once a line-of-sight deviation occurs in the next scan cycle, the corresponding ground stripe has already been acquired, and subsequent control cannot recover the spectral distortion at that moment. While the second type of method mentioned above introduces exposure time axis and disturbance spectrum analysis, its goal is not to construct rolling predictive compensation with the aim of minimizing line-of-sight stabilization error; therefore, it lacks the joint predictive ability for terrain changes and perturbation propagation over several future control cycles.

[0008] Therefore, how to construct a posture control method and system that takes the surface normal as a reference, takes the load line of sight as the control object, and can integrate the body attitude, gimbal attitude, terrain changes and vibration disturbances, and perform prediction and compensation for exposure time axis in the push-broom hyperspectral mapping process of complex terrain in mining areas has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for attitude stabilization control in hyperspectral mapping of complex terrain in mining areas, in order to solve the problems existing in the prior art, such as the compensation object deviating from the actual observation geometry, the disconnect between body control and load control, and the lag of compensation action behind the exposure time. In this way, the impact of line-of-sight mismatch and scanning direction mismatch on the quality of hyperspectral push-broom imaging is reduced in scenarios such as complex slopes, step edges, and dam slope bends.

[0010] Technical solution

[0011] To achieve the above objectives, the present invention provides the following technical solution: A method for attitude stabilization control in hyperspectral mapping of complex terrain in mining areas, applied to a flight platform equipped with a hyperspectral camera and a servo gimbal, performing load line-of-sight attitude stabilization adjustment based on multi-source sensor data. The method includes: collecting flight platform body attitude data, body angular velocity data, gimbal attitude data, terrain elevation data, ranging data, and load vibration data, and aligning the collected multi-source data with the hyperspectral camera exposure time as a unified time axis; constructing a local surface elevation window based on the aligned terrain elevation data and ranging data, calculating the terrain gradient and constructing a local surface normal vector from the local surface elevation window, and determining an ideal load line-of-sight reference based on the local surface normal vector; simultaneously solving the actual load line-of-sight based on the body attitude matrix, gimbal attitude matrix, and installation extrinsic parameter matrix, calculating the line-of-sight error angle between the actual load line-of-sight and the ideal load line-of-sight reference, and calculating the scanning direction mismatch by combining the flight platform ground velocity direction and local contour line tangent. An enhanced state vector is constructed based on at least attitude information, error state information, terrain information, relative ground height information, and vibration feature information extracted from load vibration data. A discrete prediction model is established, including the current error state term, terrain gradient change term, and vibration disturbance feature term. Within a preset prediction time domain, the gimbal compensation control sequence is solved in a rolling manner according to the discrete prediction model. The optimization objective of the gimbal compensation control sequence simultaneously constrains the line-of-sight error, line-of-sight error change rate, scan direction mismatch, control energy, control smoothness, and mechanical boundary risk. The optimal control initial value of the current cycle is superimposed with the fast feedback compensation value to generate a gimbal compensation command. The gimbal compensation command is sent to the gimbal to perform attitude stabilization compensation. When the normal reference fails, the prediction solution is abnormal, the encoder is abnormal, or the continuous prediction error exceeds the preset mismatch condition, the system switches to safety degradation control. After executing the current control cycle, new line-of-sight error, gimbal response status, and exposure synchronization information are read, the enhanced state vector for the next cycle is updated, and the state transition data of the most recent control cycle is cached for offline parameter recalibration after the flight ends. The offline parameter recalibration is used to update terrain change propagation parameters, vibration propagation parameters, and feedback gain parameters, and the parameter versions before and after the update are compared through independent verification logs. If the preset improvement conditions are not met, parameter rollback is performed.

[0012] Furthermore, the step of aligning the acquired multi-source data with the exposure time of the hyperspectral camera as a unified time axis includes: mapping flight control attitude data and gimbal attitude data to the exposure timestamp; extracting spectral features from the load vibration data within a preset control period; performing local window clipping on the digital elevation model based on the current position projection; and maintaining the most recent valid value or writing an invalid flag when the ranging data fails, based on the failure duration period.

[0013] Further, the step of constructing a local surface elevation window based on aligned terrain elevation data and ranging data, and calculating the terrain gradient and constructing a local surface normal vector from the local surface elevation window, includes: determining the local elevation construction method among ranging correction of the digital elevation model, prioritizing the use of the digital elevation model, using ranging to back-calculate the center elevation of the window, and weighted fusion of the two, based on the validity of the ranging data and its deviation relationship with the slope distance calculated by the digital elevation model; performing outlier removal on abnormal ranging values; performing difference calculations in two horizontal orthogonal directions based on the local surface elevation window to obtain the terrain gradient; constructing an unnormalized normal vector based on the terrain gradient and performing normalization processing to obtain the local surface normal vector; and obtaining the terrain gradient by using one-sided difference or preserving the nearest valid value when the elevation window is located at a missing boundary resulting in insufficient effective grid.

[0014] Furthermore, the step of determining the ideal load line-of-sight reference based on the local surface normal vector and calculating the actual load line-of-sight based on the aircraft attitude matrix, gimbal attitude matrix, and installation extrinsic parameter matrix includes: In near-normal mapping mode, determining the reverse of the local surface normal vector as the ideal load line-of-sight reference; In mapping mode that takes forward-looking safety into account, determining the ideal load line-of-sight reference based on the reverse of the local surface normal vector, the tangential direction of the flight path, and the forward safety distance; Transforming the nominal optical axis of the load to the navigation coordinate system according to the product relationship of the aircraft attitude matrix, gimbal attitude matrix, and installation extrinsic parameter matrix to obtain the actual load line-of-sight; Determining the line-of-sight error angle based on the angle between the actual load line-of-sight and the ideal load line-of-sight reference, and determining the scanning direction mismatch based on the angle between the horizontal projection direction of the flight platform ground speed and the tangential direction of the local contour line; When the terrain slope is less than a preset slope threshold, reducing the weight of the scanning direction mismatch in the control or setting it to a low-impact state.

[0015] Furthermore, the vibration features extracted from the load vibration data include a vibration coupling index and frequency-band normalized spectral energy, wherein: the vibration coupling index is calculated from the normalized correlation strength between the load base vibration acceleration spectrum and the body angular velocity spectrum; the frequency-band normalized spectral energy includes the normalized spectral energy of the lower vibration frequency band and the normalized spectral energy of the higher vibration frequency band; the rate of change of the line-of-sight error is estimated by the difference between the line-of-sight error angle of the current cycle and the smoothed line-of-sight error angle of the previous cycle; the enhanced state vector also includes state markers used to characterize the validity of the spectral sample or to characterize the vibration features following the previous cycle.

[0016] Furthermore, the establishment of a discrete prediction model comprising a current error state term, a terrain gradient change term, and a vibration disturbance feature term includes: defining the prediction state as an error state vector composed of the line-of-sight error angle, the line-of-sight error change rate, and the scanning direction mismatch; using a state evolution matrix to characterize the natural evolution of the current error state on the next cycle error state; using a control input matrix to characterize the influence of gimbal roll compensation commands and gimbal pitch compensation commands on the next cycle error state; using a terrain change propagation matrix to characterize the propagation influence of future terrain gradient changes on the next cycle error state, and using a vibration propagation matrix to characterize the propagation influence of vibration disturbance features on the next cycle error state.

[0017] Furthermore, the step of rolling the solution of the gimbal compensation control sequence within the preset prediction time domain includes: extrapolating the terrain gradient of future path points within the prediction window based on the current position, aircraft heading, ground speed vector, and local terrain data to form a future terrain gradient sequence; constructing a rolling optimization objective function using a line-of-sight error term, a line-of-sight error rate of change term, a scan direction mismatch term, a control energy term, a control smoothing term, and a risk term; solving the rolling optimization objective function using a constrained solver to obtain the gimbal compensation control sequence corresponding to multiple future control cycles; executing only the current optimal control initial value in the gimbal compensation control sequence, and re-executing the prediction and solution in the next control cycle; switching to a simplified solution mode when the processor load rate exceeds a preset load rate threshold or a solution timeout trend occurs.

[0018] Furthermore, the risk item includes at least one of the following characterizing constraints: the risk that the future line-of-sight error will exceed a preset safety error threshold; the risk that the future scanning direction mismatch will exceed a preset mismatch safety threshold; the risk that the future gimbal pitch angle will approach the mechanical limit; the risk that the future gimbal roll angle will approach the mechanical limit; wherein, the angle, angular velocity, and angular position quantities in the optimization calculation are uniformly expressed in radians.

[0019] Furthermore, the step of superimposing the optimal control initial value of the current cycle with the fast feedback compensation value to generate a gimbal compensation command, and switching to safe degradation control in case of an anomaly, includes: calculating the fast feedback compensation value based on the current line-of-sight error angle and the rate of change of the line-of-sight error, and superimposing the fast feedback compensation value with the optimal control initial value of the current cycle to obtain the final gimbal compensation command; applying angular velocity limiting to the final gimbal compensation command, and verifying the mechanical limit of the gimbal angular position based on the angular velocity integral result; when the normal reference is continuously invalid, the prediction solution continuously times out, the gimbal encoder is abnormal, or the continuous prediction line-of-sight error exceeds a preset error threshold, closing the terrain prediction term and the vibration propagation term, and retaining only the conservative feedback control based on the current error; when the ranging data and terrain elevation data fail simultaneously, switching the ideal load line-of-sight reference to the direction directly below the aircraft body, and marking the quality of the corresponding exposure line.

[0020] This invention also provides a hyperspectral mapping attitude stabilization control system for complex terrain in mining areas, comprising: a multi-source acquisition module for acquiring aircraft attitude data, aircraft angular velocity data, gimbal attitude data, terrain elevation data, ranging data, and load vibration data, and aligning the acquired multi-source data with the hyperspectral camera exposure time as a unified time axis; a terrain normal construction module for constructing a local surface elevation window based on the aligned terrain elevation data and ranging data, calculating the terrain gradient from the local surface elevation window and constructing a local surface normal vector, and determining an ideal load line-of-sight reference based on the local surface normal vector; a line-of-sight error generation module for solving the actual load line-of-sight based on the aircraft attitude matrix, gimbal attitude matrix, and installation extrinsic parameter matrix, calculating the line-of-sight error angle and scanning direction mismatch, and constructing an enhanced state vector containing vibration characteristics; and a prediction compensation module for solving the gimbal compensation control sequence in a preset prediction time domain based on a discrete prediction model containing current error state terms, terrain gradient change terms, and vibration disturbance characteristic terms, and generating gimbal compensation commands. The Anomaly Degradation and Quality Management module is used to switch to safe degradation control when the normal reference fails, the prediction solution is abnormal, the encoder is abnormal, or the continuous prediction error exceeds the preset mismatch condition, and outputs quality label information when the ranging data and terrain elevation data fail simultaneously; The Closed-Loop Parameter Update module is used to read the new line-of-sight error, gimbal response status and exposure synchronization information after the current control cycle is executed to update the enhanced state vector, cache state transition data, and perform offline parameter updates and version rollback management according to the task log.

[0021] Among them, the task security level, integrity check code, allowed execution window and prediction effective timestamp in the control instruction frame are preferred engineering implementation methods, which are used to improve the timing traceability and transmission integrity of the control link, and are not necessary limiting conditions for implementing the core attitude control method of this invention; switching to simplified solution mode when the processor load rate exceeds the preset load rate threshold is also a preferred implementation method, which is used to improve the determinism and real-time performance in engineering deployment, and is not a necessary limiting condition of this invention.

[0022] Beneficial effects

[0023] Compared with existing technologies, this invention focuses on the attitude stabilization requirements of hyperspectral push-broom imaging under complex terrain in mining areas. It forms a control process that uses the surface normal as a reference, the actual line of sight of the load as the control object, and prediction compensation, abnormal degradation, and closed-loop parameter updates as execution means, and has the following beneficial effects.

[0024] Firstly, this invention directly controls the geometric relationship between the line of sight and the surface normal, rather than compensating solely for image plane offset within the camera. By constructing a local terrain gradient and surface normal reference during the acquisition phase, defining the ideal line of sight as the reverse of the surface normal, and using the angle between the actual load line of sight and the ideal line of sight as a direct control quantity, the control target can be extended from the in-flight imaging link to the geometry of Earth observation itself. This reduces the line of sight slant error caused by complex slopes, abrupt step changes, and lateral terrain, and reduces the risk of strip geometric miscut and cross-feature spectral mixing. Corresponding test results show that in a scenario with abrupt slope changes and steps, the average line of sight error decreased from 1.84° in the control scheme to 0.76°; in actual flight tests, the average line of sight error decreased from 1.58°±0.29° to 0.93°±0.17°, and the 95th percentile line of sight error decreased from 2.74°±0.36° to 1.52°±0.21°.

[0025] Secondly, this invention unifies the aircraft attitude, gimbal angle position, terrain gradient, and vibration disturbance characteristics into the same enhanced state space, solving the problem of residual phase difference between aircraft stability and load stability. By introducing the vibration coupling index and normalized spectral energy of different frequency bands, the prediction model not only considers the current attitude error but also the error propagation trend caused by structural resonance and installation flexibility. Therefore, it can collaboratively compensate for the residual relative error between the aircraft and the load that has not been eliminated by the flight control closed loop, provided that the aircraft has been basically stabilized by the flight control. This mechanism enables the system to maintain high attitude stability continuity even when crosswinds, base vibrations, and short-term aerodynamic disturbances are superimposed. Corresponding test results show that under the condition of crosswind superimposed base vibration, the standard deviation of the lateral drift of the scan line ground projection decreased from 0.48 m to 0.19 m; in the semi-physical bench test, when superimposed with a 40 Hz sweep frequency disturbance, the peak-to-peak value of the residual oscillation of the load line decreased from 0.91° to 0.43°.

[0026] Third, this invention incorporates the scan direction mismatch as an independent state variable and cost term into the rolling optimization process, rather than compensating only for attitude angle errors. By simultaneously constraining the line-of-sight error and the scan direction mismatch, the flight platform can maintain ground-normal observation while ensuring consistency between the strip scan direction and the tangent of local contour lines, thereby improving strip orthogonality and ground projection velocity consistency. This mechanism helps suppress uneven inter-row overlap, strip edge distortion, and stitching errors caused by strip direction deviation in pushbroom imaging. Corresponding test results show that the average strip orthogonality deviation decreased from 5.8° to 2.4°, and the standard deviation of inter-row overlap deviation decreased from 0.87 pixels to 0.39 pixels; in actual flight tests, the average ground projection velocity consistency index decreased from 0.14 to 0.07.

[0027] Fourth, this invention establishes a short-term rolling predictive control mechanism that includes terrain change, vibration propagation, and mechanical boundary risk terms along the exposure time axis, transforming compensation actions from post-observation correction to pre-emptive avoidance. Since terrain gradient changes and vibration propagation trends over several control cycles are incorporated into the optimization objective in advance, the system can generate mechanically boundary-constrained compensation commands before actually scanning slope bends, curved flight segments, or areas of heightened disturbance, thereby reducing data degradation that cannot be reversed by post-event geometric correction. Simultaneously, this invention also includes anomaly degradation, quality marking, and parameter rollback mechanisms, enabling the system to still output usable control results, traceable data quality information, and verifiable parameter version management results even when sensors fail, solve timeouts occur, or encoders malfunction. Corresponding test results show that in the tailings dam slope curved flight scenario, the average strip splicing overlap error decreased from 1.26 pixels to 0.58 pixels; in actual flight tests, the effective spectral pixel percentage increased from 92.1%±1.4% to 96.2%±0.9%. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments are briefly described below. The accompanying drawings are flowcharts strongly related to the method flow of the present invention, used to illustrate the processing relationships between multi-source data acquisition, terrain normal construction, line-of-sight error generation, prediction compensation, and anomaly degradation and feedback updates.

[0029] Figure 1 The flowchart of the overall method in the embodiment of the present invention shows the overall execution flow of steps S1 to S5, including multi-source data acquisition and time alignment, terrain normal construction and enhanced state generation, predictive gimbal collaborative compensation, abnormal degradation judgment and safe execution, and closed-loop feedback and parameter update.

[0030] Figure 2The flowchart of the data acquisition and terrain normal construction unit in this embodiment of the invention shows the processing steps S101 to S103 and S201 to S202, including multi-source raw data acquisition, resampling and alignment according to the exposure time axis, distance anomaly handling, local DEM window clipping, terrain gradient calculation, surface normal construction, and ideal line-of-sight reference generation.

[0031] Figure 3 The flowchart of the line-of-sight error construction and enhancement state generation unit in this embodiment of the invention shows the processing steps S203 to S205, including actual load line-of-sight calculation, line-of-sight error and scanning direction mismatch calculation, vibration frequency band energy extraction, vibration coupling index generation, and enhancement state vector construction.

[0032] Figure 4 The flowchart of the predictive gimbal collaborative compensation and safe projection unit in the embodiment of the present invention shows the first-level judgment process of steps S301 to S305 and step S401, including error prediction model establishment, future terrain gradient extrapolation, rolling cost calculation, optimal control sequence solution, fast feedback compensation superposition, and judgment of control command overrun or solution timeout.

[0033] Figure 5 The flowchart of the abnormal degradation and feedback update unit in this embodiment of the invention shows the processing of steps S401 to S403 and steps S501 to S503, including abnormal condition judgment, degradation control execution, downward view axis switching and exposure line number recording, new error feedback reading, state transition data caching, and offline parameter update and version rollback judgment. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0035] Terminology Definition

[0036] 1. Line of sight error angle: The angle between the actual load line of sight and the ideal load line of sight reference, used to characterize the degree to which the load observation direction deviates from the optimal ground normal, in radians.

[0037] 2. Scan direction mismatch: The angle between the horizontal projection direction of the flight platform's ground velocity and the tangent of the local contour line, used to characterize the degree of matching between the direction of the push-broom strip and the main trend of the terrain, in radians.

[0038] 3. Enhanced State Vector: A multi-dimensional state vector that integrates attitude information, error state information, terrain information, relative ground height information, and vibration characteristic information is used as input to the discrete prediction model.

[0039] 4. Vibration Coupling Index: The normalized correlation strength between the vibration acceleration spectrum of the load base and the angular velocity spectrum of the body, used to characterize the strength of error propagation caused by structural coupling between the body and the load.

[0040] 5. Safety degradation control: When the normal reference fails, the prediction solution is abnormal, or the error exceeds the limit, the terrain prediction and vibration propagation terms are turned off, and only the degraded operation mode based on conservative feedback control based on the current error is retained.

[0041] 6. Discrete Prediction Model: A discrete-time state-space model that uses the current enhanced state vector as input to predict the error state of the next cycle. It includes four types of transmission relationships: error evolution, control input, terrain change, and vibration disturbance.

[0042] Example 1: Overall Flowchart of a Hyperspectral Mapping Attitude Stabilization Control Method and System for Complex Topography in Mining Areas

[0043] As shown in Figures 1 to 5, the attitude stabilization control method and system for hyperspectral mapping of complex terrain in mining areas provided in this embodiment includes five collaborative stages: multi-source data acquisition and time alignment, terrain normal construction and enhanced state generation, predictive gimbal collaborative compensation, anomaly degradation judgment and safe execution, and closed-loop feedback and parameter update. The execution entity of this method is a computer device installed on an unmanned aerial vehicle (UAV) platform. This computer device includes a processor, memory, flight control data interface, inertial measurement unit (IMU) interface, elevation or ranging interface, vibration sensor interface, gimbal encoder interface, and gimbal servo drive interface. The processor reads the control program from the memory and executes the following steps according to a fixed control cycle.

[0044] Step S1: Multi-source data acquisition and unified time alignment. This step corresponds to the data acquisition and terrain normal construction unit shown in Figure 2.

[0045] S101. Collect raw data from multiple sources. This sub-step is executed by the onboard processor.

[0046] The processor reads the aircraft attitude angle, angular velocity, flight speed, and flight altitude through the flight control data interface; the gimbal pitch angle, roll angle, and yaw compensation angle through the gimbal encoder interface; the three-axis angular velocity and three-axis acceleration through the inertial measurement unit; the vibration acceleration of the load mounting base in three directions through vibration sensors; the distance value near the center of the field of view through the airborne ranging sensor or lidar; and the elevation grid of the current location's neighborhood through a pre-loaded digital elevation model of the mining area in memory. Specifically, the aircraft attitude data characterizes the orientation state of the flight platform relative to the navigation coordinate system; the aircraft angular velocity data characterizes the current attitude change trend; the gimbal attitude data characterizes the active compensation state of the load relative to the aircraft; terrain elevation data and ranging data are used together to construct the current local surface shape; and the load vibration data characterizes the structural coupling disturbance intensity between the aircraft and the load. After this processing, the system can obtain the raw observations required for subsequent state fusion within the same control cycle.

[0047] S102. Perform resampling and time-scale alignment according to the exposure time axis.

[0048] This sub-step is still executed by the onboard processor. The processor uses the exposure time of the hyperspectral camera as a unified reference time axis, mapping data from different sampling frequencies to the same exposure timestamp. For flight control attitude and gimbal angle, the processor uses linear interpolation to map to the exposure time; for continuous angle quantities, a first-order hold method is used to ensure continuity between adjacent cycles; for discrete terrain category information, nearest neighbor hold is used. Vibration signals are first acquired at a preset sampling rate, and then spectral features are extracted within each control cycle. The digital elevation model performs local window clipping based on the current aircraft position projection. In one specific implementation, the vibration signal sampling rate can be 200 Hz, derived from bench sweep frequency testing of the aircraft-gimbal coupling frequency band, which can cover the vibration range with significant structural coupling. The control cycle can be 20 ms, and the exposure beat is synchronized with the control cycle. If the continuous loss of ranging data does not exceed a preset number of consecutive failure cycles, the nearest valid value is held; if it exceeds the preset number of consecutive failure cycles, an invalid flag is written and the process transitions to an abnormal branch. This design maps data from different sources to a uniform exposure beat scale, enabling subsequent models to predict and control based on the true geometric state corresponding to a single exposure.

[0049] S103, Perform elevation fusion, anomaly removal, and parameter dictionary loading.

[0050] This sub-step, executed by the onboard processor, addresses potential inconsistencies between the digital elevation model (DEM) and real-time ranging, and establishes a unified parameter set for subsequent control calls. The processor first compares the deviation between the ranging value and the slope distance at the center of the field of view calculated from the DEM, determining the local elevation construction method based on its validity. In one optional implementation, if the ranging value is valid and its relative deviation from the slope distance calculated by the DEM does not exceed 8%, the ranging value is used to perform a local first-order correction on the center elevation of the DEM. If the relative deviation exceeds 8% and two consecutive sampling points appear, it is determined that the ranging is affected by dust reflection or slope obstruction, and the DEM is prioritized, with an anomaly marker recorded. If the DEM is partially missing but the ranging is valid, the elevation of the ground point at the center of the field of view is calculated from the ranging value and used as a window center constraint. If both are valid but the difference is between 3% and 8%, a weighted fusion is performed. In one specific implementation, the fusion weight can be 0.7, which can be obtained through historical working condition statistics and offline calibration of elevation fitting errors.

[0051] If the deviation of a single-point distance measurement relative to two adjacent points exceeds 5 m, the distance measurement is considered an outlier and discarded. This threshold can be obtained offline based on the outlier statistics of the distance measurement sensor in the dusty environment of the mining area.

[0052] In addition, the processor loads a unified parameter dictionary. In one specific implementation, the control cycle can be 20 ms, the local window size for digital elevation modeling can be 21×21 grids with a grid side length of 2 m, the vibration spectrum analysis frequency band can be 20 Hz to 80 Hz, the upper limits for gimbal pitch and roll compensation angles can be ±15° and ±12° respectively, the upper limit for single-cycle angular velocity commands can be 25° / s, the preset safety error threshold can be 1.2°, and the preset error threshold can be 2.5°. These preset parameters can be obtained offline based on the target gimbal's execution capabilities, DEM resolution, and statistical results from multiple flight tests. For optimization calculations, all angle, angular velocity, and angular position quantities are uniformly converted to radians to ensure dimensional consistency.

[0053] In one specific implementation, the refresh rate of the dual-axis gimbal encoder can be 500 Hz, and the refresh rate of the servo drive command can be 250 Hz. Bench step tests show that the response time from 10% to 90% of the pitch axis is 14.6 ms, and the response time from 10% to 90% of the roll axis is 15.8 ms, with equivalent closed-loop bandwidths of 12.4 Hz and 11.8 Hz, respectively. With a 20 ms control cycle, the solver's average timeout is 4.3 ms, the 95th percentile timeout is 6.9 ms, and the maximum timeout is 7.6 ms. Therefore, 8 ms can be used as the solver timeout threshold. This threshold is determined based on the upper limit of the solver's runtime log statistics and with control redundancy. Switching to a simplified solver mode when the processor load exceeds 85% is a preferred engineering implementation method used to maintain real-time performance and determinism when computing power is limited, but it does not constitute a necessary condition for implementing the core technical solution of this invention. Once this parameter dictionary is loaded, subsequent terrain modeling, prediction solving, and anomaly detection can all be performed at a unified scale.

[0054] Step S2: Terrain normal construction and enhanced state generation.

[0055] S201. Construct a local elevation window and calculate the terrain gradient and surface normal vector.

[0056] This sub-step, executed by the onboard processor, is used to recover the optimal observation reference direction from the local terrain shape near the current UAV location. The processor performs differential calculations in two horizontally orthogonal directions, east and north, from the local elevation window output in step S1 to obtain the terrain gradient; then, it uses the terrain gradient to construct the local surface normal vector.

[0057] In this embodiment, the local terrain elevation function is set as follows: ,in, This indicates the horizontal coordinates facing east, in meters. This indicates the northward horizontal coordinate, in meters. Elevation is indicated in meters; This represents the surface elevation function obtained after fusing digital elevation model and distance measurement data. The processor performs central difference calculation on the local window center to obtain the eastward and northward elevation gradients:

[0058] ,in, Indicates the eastward elevation gradient; Indicates the northward elevation gradient; This indicates the step size of the elevation grid, in meters. In one specific implementation, A value of 2 m can be chosen, which is determined by the resolution of the digital elevation model grid.

[0059] The processor constructs an unnormalized normal vector based on the terrain gradient and performs normalization processing to obtain a local surface normal vector:

[0060] ,in, This represents the unnormalized normal vector; Represents the unit surface normal vector; Representing vectors The second norm; This represents a smoothing term to prevent the denominator from approaching zero. In one specific implementation, Desirable This smoothing term can be set based on the numerical stability simulation results to avoid numerical anomalies during the normalization process.

[0061] The above formula corresponds to the surface normal construction process in step S2. The input data comes from the local elevation window after distance measurement correction. The technical sub-problem it solves is to obtain a surface normal vector that can be directly used to construct the ideal observation direction from the local terrain elevation distribution. The terrain gradient output by the formula reflects the degree of inclination of the surface in two orthogonal directions, and the output unit normal vector is directly used in the generation of the ideal line-of-sight reference in the next sub-step.

[0062] When a local window is located on a boundary lacking elevation and there are insufficient effective grid cells, the processor does not use central differencing, but instead uses lateral differencing or the most recent valid value preservation. Specifically, when the number of effective grid cells is less than 9, lateral differencing of the three most recent valid points can be used to estimate the gradient; when the local terrain is approximately horizontal and the normalized result of the unnormalized normal vector is close to vertically upward, the system treats it as a natural degradation case rather than an anomaly. After this processing, subsequent reference line generation can maintain continuity even with abrupt changes in local terrain and missing elevation boundaries.

[0063] S202. Generate an ideal load line-of-sight reference based on the local surface normal vector.

[0064] This sub-step is executed by the onboard processor. Based on the current mission mode, the processor converts the local surface normal vector into an ideal load line-of-sight reference vector. For near-normal mapping mode, the reverse surface normal vector is directly used as the ideal line-of-sight reference; for mapping modes that also consider forward-looking safety, the reverse surface normal vector is weighted and fused with the track tangential direction, ensuring that the line-of-sight reference meets mapping requirements while retaining a certain amount of forward-looking information.

[0065] In one specific implementation, when the forward safe distance is greater than 0.8 times the flight altitude, a near-normal mapping mode is adopted; when the forward safe distance is less than 0.8 times but greater than 0.4 times the flight altitude, a weighted fusion of the normal direction and the tangential direction of the flight path is adopted. The fusion weight can be determined jointly based on the obstacle avoidance markers and the safe distance, and this weight can be obtained through multiple simulations and offline calibration of obstacle avoidance trigger statistics; when the forward safe distance is less than 0.4 times the flight altitude, obstacle avoidance is prioritized, and the main attitude control mode of this invention is not entered. The main implementation adopts the near-normal mapping mode.

[0066] The output of this sub-step is the ideal payload line-of-sight reference vector and elevation confidence marker. By directly anchoring the observation target to the opposite direction of the local surface normal, the subsequent control objective is no longer simply to maintain the stability of the aircraft or camera interior, but rather to revolve around the geometrically optimal direction for Earth observation.

[0067] S203. Calculate the actual load line of sight based on the aircraft attitude, gimbal attitude, and installation external parameters.

[0068] This sub-step, executed by the onboard processor, is used to determine the true orientation of the load's optical axis in the navigation coordinate system. The processor transforms the load's nominal optical axis to the navigation coordinate system according to the product relationship between the airframe attitude matrix, the gimbal attitude matrix, and the installation extrinsic parameter matrix.

[0069] In this embodiment, the navigation coordinate system adopts the Northeast-Sky coordinate system, denoted as . System, among which The axis points east. The axis points north. The axis points upwards; the body coordinate system is denoted as... System, among which The axis points towards the machine head. The axis points to the right side of the fuselage. The axis points downwards from the body; the gimbal coordinate system is denoted as... The zero point of the load coordinate system is parallel to the body coordinate system; the load coordinate system is denoted as... The nominal direction of the load optical axis is... .

[0070] The aircraft attitude is constructed using a rotation matrix based on the sequence of yaw, pitch, and roll.

[0071] ,in, Represents the body attitude matrix; Indicates the aircraft's yaw angle, expressed in radians; Indicates the aircraft's pitch angle, expressed in radians; Indicates the roll angle of the aircraft, in radians; , and These represent the standard right-handed rotation matrices about the corresponding coordinate axes.

[0072] For a dual-axis gimbal, the mechanical structure includes a sequential axis arrangement of roll followed by pitch, and its attitude matrix can be represented as:

[0073] ,in, This represents the gimbal attitude matrix; This indicates the gimbal's pitch angle, expressed in radians. This indicates the roll angle of the gimbal, expressed in radians. If the platform is configured as a three-axis gimbal with a yaw compensation mechanism, the yaw compensation angle can be expanded to include this feature. However, in the main protection method of this embodiment, a two-axis gimbal structure is used.

[0074] The extrinsic parameter matrix is ​​used to compensate for the fixed deflection angle of the load relative to the gimbal mounting surface. Its calibration form can be expressed as:

[0075] ,in, This indicates the installation extrinsic parameter matrix; , and These represent the installation yaw angle, installation pitch angle, and installation roll angle, respectively, all in radians. The installation micro-angles can be obtained offline through static target calibration and ground attitude tests.

[0076] Therefore, the actual load line vector can be expressed as:

[0077] ,in, This represents the actual load unit vector along the line of sight; This indicates the nominal orientation of the load optical axis in its body coordinate system; , and All are 3×3 rotation matrices.

[0078] The above formula corresponds to the actual line-of-sight calculation process in step S2. The input data comes from the flight control attitude fusion results, gimbal encoder calculation results, and installation calibration parameters, respectively. The technical sub-problem it solves is to map the large-scale orientation of the airframe, the active compensation attitude of the gimbal, and the mechanical installation deviations to the same navigation coordinate system, and output a line-of-sight vector that can accurately represent the direction of Earth observation. This output will be directly used for subsequent line-of-sight error calculation.

[0079] When attitude data is temporarily lost, the processor uses the matrix from the previous cycle and adds an expired timestamp. If packet loss continues for more than three cycles, predictive control stops and the system enters a conservative hold mode. This design allows the system to output continuous line-of-sight status even under short-term communication jitter.

[0080] S204. Calculate the mismatch between the viewing axis error angle and the scanning direction.

[0081] This sub-step, executed by the onboard processor, quantifies the degree to which the current observation state deviates from the ideal state geometrically. The processor first calculates the line-of-sight error angle based on the angle between the actual load line-of-sight vector and the ideal load line-of-sight reference vector:

[0082] ,in, This represents the angular error between the actual line of sight and the ideal reference, expressed in radians. This represents the dot product of the actual load line-of-sight vector and the ideal load line-of-sight reference vector. This indicates a clipping operation, used to restrict the dot product result to the domain of the inverse cosine function, avoiding undefined calculations caused by floating-point errors; This represents the ideal load line-of-sight reference vector.

[0083] Subsequently, the processor calculates the scan direction mismatch based on the angle between the horizontal projection direction of the flight platform's ground speed and the tangent of the local contour lines:

[0084] ,in, This indicates the angular deviation between the scanning direction and the tangent of the local contour line, expressed in radians. The unit vector representing the horizontal projection of the ground velocity vector; A unit vector representing the tangent of a local contour line; It is obtained by normalizing the projection of the body's ground velocity vector onto the horizontal plane; It is obtained by normalizing the horizontal orthogonal direction of the terrain gradient.

[0085] The above two equations correspond to the error state generation process in step S2. The input of the first equation is the actual load line of sight and the ideal line of sight reference, and the output is the line of sight angle error directly used for attitude control; the input of the second equation is the horizontal movement direction of the aircraft and the tangential direction of the local terrain, and the output is the scanning direction mismatch reflecting the consistency between the strip scanning direction and the main direction of the terrain. Both are then fed into the subsequent enhancement state vector and prediction cost function.

[0086] To reflect the consistency of ground projection velocity, the processor can also calculate the scanning normal projection velocity deviation as an auxiliary diagnostic measure. In one optional implementation, when the local slope is very small and the tangential modulus of the local slope contour lines is less than 0.05 (slope threshold 0.05), it indicates that the terrain is close to horizontal, and the impact of scanning direction mismatch on spectral quality is reduced. The processor sets the scanning direction mismatch to 0 and reduces its weight in the cost function. This threshold can be obtained through offline calibration based on near-horizontal surface simulation results.

[0087] When the normal reference is invalid, the processor sets the line-of-sight error angle to the previous valid value and adds a weight reduction flag; when the local velocity vector or local contour tangent is invalid, the scan direction mismatch is set to 0 and a low confidence flag is added.

[0088] S205. Extract vibration features and construct an enhanced state vector.

[0089] This sub-step is executed by the onboard processor. The processor uniformly writes the aircraft attitude, aircraft angular velocity, gimbal angular position, line-of-sight error angle, line-of-sight error rate of change, scan direction mismatch, terrain gradient, relative ground height, and vibration characteristics into the enhanced state vector.

[0090] In this embodiment, the first The enhanced state vector for each control cycle can be defined as:

[0091] ,in, Indicates the first The enhanced state vector for each control cycle; , and These represent the aircraft's roll angle, pitch angle, and yaw angle, respectively, in radians; , and These represent the three-axis angular velocities of the machine system, in radians per second. , and These represent the gimbal roll angle, pitch angle, and yaw compensation angle, respectively, in radians; This indicates the line-of-sight error angle, in radians. This represents the rate of change of the line-of-sight error, expressed in radians per second. Indicates the mismatch in the scanning direction, in radians; and Indicates terrain gradient; This indicates the height relative to the ground, expressed in meters. Indicates the vibration coupling index; and These represent the normalized spectral energies for the lower and higher vibrational frequency bands, respectively.

[0092] The vibration coupling index is used to characterize the normalized correlation strength between the vibration acceleration spectrum of the load base and the angular velocity spectrum of the body. Its calculation expression is as follows:

[0093] ,in, Indicates the vibration coupling index; Represents frequency The amplitude spectrum of vibration acceleration at a given location; Represents frequency The angular velocity amplitude spectrum at that location; This represents a smoothing term to prevent the denominator from being zero. In one specific implementation, Desirable The value is determined by numerical stability requirements. Since a non-negative amplitude spectrum is used, this index can characterize the relative strength change from low to high coupling. This state variable is defined as an estimate of the disturbance propagation intensity caused by structural coupling between the body and the load; its source is historical short-window data from vibration sensors and body angular velocities; it is obtained through spectral analysis and normalized correlation calculations; its rationality can be verified by comparing the synchronous increase in residual oscillation of the line of sight when vibration intensifies in bench sweep frequency tests.

[0094] Normalized spectral energy by frequency band is used to describe the proportion of different vibration frequency bands in the total vibration energy. In one specific implementation, spectral feature extraction can use a sliding window with a length of 32 sampling points. A 50% overlap update is performed on a 200 Hz sampling vibration sequence corresponding to a 160 ms duration. A Hamming window function is used, with a frequency resolution of 6.25 Hz. When there are fewer than 32 points, zero-filling is used and the confidence level is reduced. The window length and overlap ratio can be selected offline based on the spectral peak stability in bench repeated tests.

[0095] The rate of change of line-of-sight error is obtained by dividing the difference between the line-of-sight error angle of the current period and the smoothed line-of-sight error angle of the previous period by the control period. In one specific implementation, the smoothing coefficient can be 0.35, which can be obtained through offline calibration by compromising between error sequence noise suppression and response speed. If the instantaneous change in line-of-sight error within a certain period exceeds three times that of the previous period and a ranging anomaly marker exists simultaneously, low-pass smoothing is performed first before updating the rate of change of line-of-sight error to avoid noise spikes directly entering the prediction model.

[0096] When the total energy of a frequency band in a certain control cycle is zero, i.e., the platform is in a near-static condition, the vibration coupling index can be directly set to 0. When there are insufficient samples within the spectrum window, the vibration characteristics of the previous cycle are used and a low-confidence flag is added. The low-confidence flag or the used state flag is used to characterize the validity of the spectrum samples and whether the current vibration characteristics are inherited from the previous cycle. After this processing, the enhanced state vector can simultaneously characterize the current geometric error, terrain change trend, and vibration propagation risk, providing a unified input for subsequent predictive control.

[0097] Step S3: Predictive gimbal collaborative compensation.

[0098] S301. Establish a discrete error prediction model.

[0099] This sub-step, executed by the onboard processor, predicts the residual error state for the next cycle based on the current error state, terrain changes, and vibration disturbances. The processor defines the predicted state as an error state vector consisting of the line-of-sight error angle, the rate of change of the line-of-sight error, and the scan direction mismatch, and establishes a discrete state update relationship:

[0100] ,in, Indicates the first The periodic error state vector; This represents the error state vector for the next period. This represents the natural evolution matrix of the error, used to characterize the natural propagation of the current error state to the error state of the next period; This represents the control input matrix, used to characterize the impact of gimbal compensation control on the error state of the next cycle; This indicates the gimbal roll compensation angular velocity command and the gimbal pitch compensation angular velocity command for the current cycle, in radians per second. Indicates the amount of change in terrain gradient; This represents the terrain change propagation matrix, used to characterize the propagation effect of future terrain gradient changes on the error state; Represents the characteristic vector of vibration disturbance; This represents the vibration propagation matrix, used to characterize the propagation effect of vibration disturbance characteristics on the error state.

[0101] The above formula corresponds to the core prediction process in step S3. The input data comes from the enhanced state vector, future terrain change estimation, and vibration characteristics. The technical sub-problem it solves is to integrate the current residual error, the terrain change ahead, and the structural coupling disturbance into the error prediction of the next cycle. The error state of the next cycle output by the formula will further participate in the rolling optimization.

[0102] In one specific implementation, the control period can be 20 ms. The error natural evolution matrix, control input matrix, terrain change propagation matrix, and vibration propagation matrix can all be obtained through least-squares identification of flight logs. The initial values ​​of the matrix elements can be calibrated using bench sweep tests and standard maneuver flight data, and then corrected using actual flight logs. To prevent excessive predictions caused by excessively large matrix elements, a constraint that the absolute value of each element does not exceed 0.35 can be added during identification. When the terrain gradient cannot be updated, the terrain gradient change is set to a zero vector; when the vibration features are invalid, the vibration disturbance feature vector is set to a zero vector, causing the model to degenerate into a normal error prediction model without terrain feedforward and vibration propagation; when the control period jitter does not exceed ±2 ms, the actual period can be absorbed into the time-related term of the natural evolution matrix.

[0103] S302, extrapolation prediction of future terrain gradient sequence within the window.

[0104] This sub-step, executed by the onboard processor, provides short-term terrain change information ahead for rolling optimization. Based on the current position, aircraft heading, ground speed vector, and local digital elevation model (DEM), the processor advances the spatial location of future path points within the prediction window and queries the corresponding terrain gradient from the DEM to form a future terrain gradient sequence.

[0105] In one specific implementation, the prediction time domain length can be taken as 5 control cycles, or 100 ms. This time domain can be determined offline based on the gimbal response time, controller solution time, and the rate of change of terrain in the mining area. If a future path point crosses the boundary of the digital elevation model, the boundary point gradient is used with an additional conservative penalty flag. In turning scenarios, path advancement uses tangential updates extrapolated from the current yaw angle and angular velocity to avoid querying the terrain ahead based on a straight path. This sub-step outputs a sequence of terrain gradient changes for multiple future control cycles and feeds it into the rolling optimization process of the next sub-step.

[0106] S303. Construct a rolling optimization objective and solve for the gimbal compensation control sequence.

[0107] This sub-step is executed by the onboard processor. Within a preset prediction time domain, the processor constructs a rolling optimization objective function using line-of-sight error, line-of-sight error change rate, scan direction mismatch, control energy, control smoothness, and a risk term. The risk term is used to preemptively penalize situations where future line-of-sight error exceeds a threshold, future scan direction mismatch exceeds a threshold, or the gimbal pitch or roll angle approaches its mechanical limits.

[0108] In one specific implementation, the prediction time domain can be 5 periods, and the decision variable dimension is 10; the control weights can be chosen as follows: , The scanning direction mismatch weight can be taken as follows: The weights for control energy (0.15), control smoothness (0.12), and risk (1.5) can be set. The risk weights for scanning direction (0.7), pitch mechanical boundary risk (1.2), and roll mechanical boundary risk (1.0) can be set offline based on simulation scan results and bench experiments, balancing line-of-sight stability, strip orthogonality, and gimbal lifespan. The safe line-of-sight error threshold can be set to 1.2°, and the safe threshold for scanning direction mismatch can be set to 4.0°, both converted to radians during optimization calculations. The gimbal pitch and roll mechanical limits can be set to 15° and 12° respectively, also converted to radians during optimization.

[0109] The inputs to this sub-step are the current error state, the future terrain gradient sequence, the current gimbal angular position, and the parameter dictionary. The output is the gimbal compensation control sequence corresponding to multiple future control cycles. The solver can preferably be a constrained quadratic programming solver, with constraints including angular velocity limiting constraints and angular position limiting constraints obtained from the integral of angular velocity. When the processor load exceeds 85% or there has been a timeout in the previous cycle, it can automatically switch to a simplified solution mode to improve determinism. This switching is a preferred engineering implementation method, its purpose being to ensure real-time solution stability, and is not a core limitation that this invention must rely on. If the local optimization solution is not completed within 8 ms, the optimization for this cycle is terminated, and the first increment of the solution from the previous cycle plus conservative decay is used for execution. This timeout threshold can be set offline based on the statistical upper limit of the solution time on the target processor.

[0110] S304. Perform rolling optimization of the initial quantity and generate prediction output.

[0111] This sub-step is executed by the onboard processor. The processor, based on the control sequence for multiple future control cycles obtained in step S303, executes only the optimal control initial value for the current cycle, and then re-predicts and solves for the next control cycle, forming rolling control. This strategy avoids the direct accumulation of long-term prediction errors at the execution end, while enabling the controller to continuously use newly acquired exposure synchronization data to correct subsequent decisions.

[0112] If the solver returns an infeasible solution, it outputs a conservative control flag and proceeds to the anomalous degradation process. The output of this sub-step includes the current period's optimal control initial value and the next period's predicted state, providing input for rapid feedback superposition and anomaly detection.

[0113] S305, superimpose rapid feedback compensation and generate the final gimbal compensation command.

[0114] This sub-step is executed by the onboard processor. The processor calculates the fast feedback compensation amount based on the current line-of-sight error angle and the rate of change of the line-of-sight error, and adds it to the optimal control initial value for the current cycle to obtain the final gimbal compensation command:

[0115] ,in, Indicates the first Rapid feedback compensation for incremental changes during the cycle; Represents the feedback gain matrix; This indicates the current cycle line error angle, in radians. This represents the rate of change of the current period's line-of-sight error, expressed in radians per second. This formula corresponds to the rapid residual compensation process in step S3. The input is the current residual error state, and the technical sub-problem it addresses is the rapid suppression of short-term residuals that are difficult to cover in a timely manner by rolling optimization. The output feedback compensation amount is superimposed with the feedforward optimization initial amount to form the final instruction.

[0116] In one specific implementation, the feedback gain matrix can be expressed as:

[0117] ,

[0118] The first row corresponds to the roll axis feedback gain, and the second row corresponds to the pitch axis feedback gain. Each element of the matrix can be obtained through joint calibration using flight logs and the gimbal's step response. The values ​​described above are only selectable values ​​in a specific embodiment and can be adjusted offline according to different gimbal response characteristics.

[0119] Subsequently, the processor applies angular velocity limiting to the final gimbal compensation command and verifies the mechanical limit of the gimbal angular position based on the angular velocity integral result. If the superimposed command exceeds the single-cycle angular velocity limit, it is directly limited to a preset limit; in one specific implementation, this limit can be ±25° / s, determined based on the safe speed range of the gimbal servo driver and bench test results. Angular velocity limiting is executed before angular position integration. If the predicted angular position obtained by integration is still close to the mechanical limit, it is constrained by increasing the risk cost and performing a secondary limit verification at the end of the current cycle. If the solution result or feedback result is non-numerical, the optimization result is discarded, and only the zero command or the conservative command from the previous cycle is retained. After this processing, the output gimbal compensation command can take into account prediction feedforward, error residual suppression, and mechanical safety boundaries.

[0120] In one optional implementation, the gimbal compensation command can be encapsulated into a control command frame and sent to the gimbal execution terminal. The control command frame may include a command number, a controlled device number, a generation timestamp, a predicted effective timestamp, a discrete compensation step size, a roll target angular velocity, a pitch target angular velocity, a mission safety level, and an integrity check code. The gimbal execution terminal determines whether the command is within the allowed execution window based on the generation timestamp and the predicted effective timestamp, and determines whether the command has been truncated, tampered with, or out of order during transmission based on the integrity check code. The above fields are set as a preferred engineering implementation method to ensure timing traceability and transmission integrity in the exposure synchronization control link, and are not necessary limitations for implementing the main control logic of this invention.

[0121] Step S4: Abnormal degradation judgment and safe execution.

[0122] S401. Determine whether degradation mode has been triggered.

[0123] This sub-step is executed by the onboard processor. The processor determines whether any of the following conditions are met: the normal reference is continuously invalid for more than a preset number of cycles, the prediction solution times out for more than a preset number of times, the gimbal encoder is malfunctioning, or the prediction line-of-sight error exceeds a preset threshold for multiple consecutive cycles. If not, the main control mode is maintained; if so, the degradation control process is entered.

[0124] In one specific implementation, the continuous invalid threshold for the normal reference can be taken as 3 cycles, the continuous timeout threshold for prediction solution can be taken as 2 cycles, and the continuous prediction line-of-sight error exceeding the threshold condition can be taken as the prediction line-of-sight error exceeding 2.5° for 3 consecutive cycles. The above thresholds can be obtained offline based on flight safety test and misjudgment rate statistics.

[0125] S402, Implement degradation control.

[0126] This sub-step is executed by the onboard processor. In degradation mode, the processor disables terrain prediction and vibration propagation terms, retaining only conservative feedback control based on the current line-of-sight error and its rate of change. Simultaneously, the processor lowers the gimbal pitch and roll rate limits and sends a speed reduction recommendation to the flight control system to mitigate coupling disturbances between the airframe and the payload.

[0127] In one specific implementation, the upper limit of angular velocity in degradation mode can be reduced to 10° / s. This upper limit is set based on the principle of prioritizing stability over rapid tracking under abnormal conditions. The cruise speed can be reduced by 20% from its original value. This percentage can be set offline based on the trade-off between flight safety and mapping continuity through multiple anomaly recovery tests. After this processing, even if local terrain references or prediction solutions are unavailable, the system can still output conservative and executable attitude control results.

[0128] S403, Perform downward view axis switching and quality marking.

[0129] This sub-step is executed by the airborne processor. If both the ranging data and the digital elevation model fail simultaneously, the processor temporarily switches the ideal load line-of-sight reference to the direction directly below the aircraft to ensure that the mapping load maintains a basic ground-view state; at the same time, it records the exposure row number corresponding to the abnormal time period, which is used to add quality markers to the data of that time period in the subsequent ground processing stage, instead of stitching it together with the normal data with equal weight.

[0130] The output of this step includes a degradation control flag, a downward-facing line switching flag, and an exposure row number index. This allows subsequent ground processing stages to identify anomalous acquisition intervals and prevents exposure data from abnormal pose conditions from being mistakenly included in normal strip stitching.

[0131] Step S5: Closed-loop feedback and parameter update.

[0132] S501, Read the new error feedback and update the status for the next cycle.

[0133] This sub-step is executed by the onboard processor. At the end of the current cycle, the processor reads the new actual load line of sight, gimbal response angle, and exposure synchronization marker, recalculates the new line of sight error angle, line of sight error rate of change, and scan direction mismatch, and updates the enhancement state vector for the next cycle. After this processing, the prediction model in step S3 can be resolved around the latest observations, thus forming closed-loop rolling control.

[0134] S502, cache state transition data.

[0135] This sub-step is executed by the onboard processor. The processor writes the state transition data from the most recent control cycles to a local cache for offline model parameter reestimation after flight. The state transition data includes at least the augmented state vector, control input, terrain gradient changes, vibration characteristics, and the error state for the next cycle.

[0136] In one specific implementation, data from the most recent 200 control cycles can be cached. The length of this window can be set offline based on local storage capacity and the minimum number of samples required for offline identification. By caching recent state transition processes, subsequent offline recalibration can adjust the propagation matrix and feedback gain for specific mining scenarios.

[0137] S503, Perform offline parameter recalibration and version rollback judgment.

[0138] This sub-step is performed at the ground station after the flight. The ground station uses the actual flight log to minimize the one-step prediction error and recalibrates the terrain change propagation parameters, vibration propagation parameters, and feedback gain parameters. The updated parameters need to be replayed and verified on an independent verification log, and the changes in indicators such as mean line-of-sight error, 95th percentile line-of-sight error, and mean scan direction mismatch before and after the update are compared.

[0139] In one specific implementation, if the average line-of-sight error does not improve, or the 95th percentile error increases by more than 5%, or the average scan direction mismatch increases by more than 8%, the system rolls back to the previous stable version. These thresholds can be set offline based on parameter update risk control requirements and historical playback statistics. This recalibration and rollback mechanism enables the system to continuously correct propagation parameters in different mining scenarios, while avoiding attitude stabilization performance degradation caused by erroneous updates.

[0140] Example 2: Airborne Hyperspectral Attitude Stabilization Control in Open-Pit Mine Bench Slope Scenarios

[0141] Open-pit mine bench slopes typically feature slope top platforms, steep embankments, turning steps, and localized mining pits. When the flight platform passes along the survey line, the surface normal can abruptly change within a short distance. Even with good attitude control, traditional fixed-attitude flight schemes can still result in stripe misalignment along the hyperspectral push-scan direction due to the load line of sight not adjusting synchronously with the slope normal. If only post-hoc geometric correction is relied upon, the already occurred cross-step mixed pixels cannot be recovered.

[0142] This embodiment utilizes a hexacopter UAV platform with a body weight of 12 kg. It is equipped with one pushbroom hyperspectral camera, one dual-axis gimbal, one inertial measurement unit, one laser rangefinder, and one base vibration sensor. The flight altitude is maintained at 120 m relative to the ground, with a cruising speed of 8 m / s. The control cycle and exposure time are fixed at 20 ms. The digital elevation model has a grid resolution of 2 m and a local window size of 21×21. The gimbal's pitch range is ±15°, and its roll range is ±12°. The onboard processor is a quad-core edge computing unit with a clock speed of 2.0 GHz; the encoder refresh rate is 500 Hz, and the servo drive refresh rate is 250 Hz.

[0143] Within a certain control cycle, the airborne processor read the following data: airframe roll angle 2.6°, airframe pitch angle -1.8°, airframe yaw angle 86.0°; airframe angular velocities 3.2° / s, -2.4° / s, and 1.1° / s; gimbal roll angle -1.2°, pitch angle 4.5°, yaw compensation angle 0.6°; relative ground altitude 118.4 m. The vibration sensor extracted normalized spectral energy of 0.34 for the lower vibration frequency band and 0.21 for the higher vibration frequency band within the most recent 160 ms sliding window. The vibration coupling index calculated from the acceleration and angular velocity spectra was 0.62. The relative deviation between the current ranging value and the slant range calculated by the digital elevation model was 4.1%, therefore, elevation fusion was performed with a weight of 0.7, which was derived from the offline calibration results of historical elevation fitting error statistics on the same platform.

[0144] Subsequently, the processor clips a local digital elevation model window from the current location, calculating an eastward gradient of 0.28 and a northward gradient of -0.16. This yields the local surface normal vector, with the ideal line-of-sight reference being the reverse of this normal vector. The actual load line-of-sight vector is obtained by solving the aircraft attitude matrix, gimbal attitude matrix, and installation calibration matrix. Further calculations show the current cycle line-of-sight error to be 0.043 rad, and the previous cycle smoothed line-of-sight error to be 0.036 rad, resulting in a current cycle line-of-sight error change rate of 0.35 rad / s. The angle between the current ground speed horizontal projection and the local contour tangent is 0.067 rad. This forms an 18-dimensional enhanced state vector, which is then fed into the predictive collaborative compensation component.

[0145] In the prediction compensation process, the prediction time domain is used for 5 cycles. Based on the current speed of 8 m / s and heading of 86°, the processor projects the future position within 100 ms ahead and extracts five gradient points from the digital elevation model to form a gradient change sequence. The optimal control increments for the current cycle are calculated to be -0.142 rad / s for roll and -0.196 rad / s for pitch; the fast feedback increments are -0.018 rad / s for roll and -0.015 rad / s for pitch. Therefore, the final executed commands are -0.160 rad / s for roll and -0.211 rad / s for pitch, which does not exceed the single-cycle angular velocity limit and does not require limiting. Corresponding to the gimbal servo, the gimbal roll angle is updated to -1.38° and the pitch angle is updated to 4.26° at the end of this cycle. The cycle time was 4.7 ms, the average servo execution delay was 2.1 ms, and the total link met the 20 ms control cycle requirement.

[0146] Within this period, the normal reference was effective, the prediction and solution time was less than 8 ms, and the continuous prediction error did not exceed the threshold, therefore, the degradation mode was not triggered. After resampling in the next cycle, the new actual line-of-sight error decreased to 1.71°, and the scan direction mismatch decreased to 2.61°. This indicates that at the edge of a step change, the system can rotate the gimbal one to two cycles in advance before actually scanning the steep step, reducing the risk of cross-step pixel mixing.

[0147] In this embodiment, the aircraft attitude and angular velocity are output by flight control and inertial navigation fusion; the gimbal angle is obtained by a high-resolution encoder; the digital elevation model is downsampled to 2 m resolution from existing topographic survey results in the mining area; the installation calibration matrix is ​​obtained by static target calibration; and the predicted propagation parameters and feedback gain are obtained by least-squares identification from 20 test flight logs. The frequency band selection of 20 Hz to 80 Hz comes from the gimbal test bench frequency sweep results, indicating that structural coupling is relatively obvious within this frequency band. The dual-axis gimbal response parameters are obtained from statistics of 50 step tests completed on the ground test bench. The average pitch axis response time is 14.6 ms with a standard deviation of 1.1 ms, and the average roll axis response time is 15.8 ms with a standard deviation of 1.3 ms.

[0148] The exemplary simulation in this embodiment uses a digital twin environment of a mining area, with a scene area of ​​1.2 km × 0.8 km and a simulation duration of 600 s. The normal operating condition is a 3 m / s crosswind, and the impact condition is an additional 0.4 s gust of 6 m / s and 40 Hz structural vibration enhancement when passing over the edge of a step. Control scheme A is fixed attitude flight with post-flight geometric correction, and control scheme B is ordinary gimbal PID stabilization without using terrain normal reference and prediction terms. The simulation was repeated 10 times, and the mean results are as follows: For control scheme A, the mean line-of-sight error was 1.84°, the 95th percentile error was 3.42°, and the mean scanning direction mismatch was 5.6°; for control scheme B, the standard deviation of the lateral drift of the scan line ground projection was 0.48 m, the limit period ratio was 8.7%, and the mean scanning direction mismatch was 4.9°; for the scheme of this invention, the mean line-of-sight error was 0.76°, the 95th percentile error was 1.31°, the standard deviation of the lateral drift was 0.19 m, the mean scanning direction mismatch was 2.1°, and the limit period ratio was 2.6%.

[0149] Twelve flight tests were conducted in a bench slope area of ​​an open-pit mine, with a total flight time of 96 minutes and 48 effective survey lines. Each survey line ranged in length from 0.8 km to 1.4 km. The control procedure employed alternating flights with the same platform and load. Six flights utilized the attitude stabilization control method of this invention, while the other six used a conventional gimbal with PID stabilization and post-flight geometric correction. Statistical results showed that the control scheme had a mean line-of-sight error of 1.58°±0.29°, a 95th percentile error of 2.74°±0.36°, a mean scan direction mismatch of 4.37°±0.58°, and a mixed pixel rate of 11.6%±1.9%. The invention's scheme had a mean line-of-sight error of 0.93°±0.17°, a 95th percentile error of 1.52°±0.21°, a mean scan direction mismatch of 2.48°±0.41°, and a mixed pixel rate of 7.4%±1.2%.

[0150] Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Example 1, completing operations such as abnormal degradation processing, closed-loop feedback, and offline parameter recalibration. Its specific execution logic is consistent with that of Example 1, and will not be repeated here.

[0151] Example 3: Hyperspectral strip stability mapping under the scenario of a tailings dam slope bend in the flight path

[0152] Tailings dam slopes typically exhibit bending boundaries, toe platforms, and localized seepage deformation zones. When drones fly near the dam crest or toe, they generate pitch and roll coupled disturbances at turns and slope transitions, leading to increased overlap errors between push-broom strips. While traditional post-exposure geometric correction can correct some attitude recording errors, it cannot recover the effective spectral pixel loss caused by rapid line-of-sight shift at the time of exposure.

[0153] This embodiment uses a quadcopter UAV platform, equipped with a hyperspectral payload similar to that in Embodiment 2. The flight altitude relative to the ground is 100 m, and the cruising speed is 6 m / s. All other control parameters remain consistent with Embodiment 2 to ensure the use of a unified parameter dictionary across different mining scenarios. The digital elevation model for the tailings dam scenario is derived from the latest laser measurement results. Due to the rapid change in the local normal in the turning area, a prediction window of 5 cycles is still used, but the turning trajectory tangential update is added to the future path point calculation to ensure that the forward-looking terrain query location is consistent with the actual flight direction.

[0154] During a control cycle before a turn, the aircraft's roll angle is -3.1°, pitch angle is 2.2°, and yaw angle is 132.4°; the gimbal's roll angle is 1.6° and pitch angle is -3.8°; the current relative ground altitude is 102.1 m; local gradient calculations show an eastward gradient of -0.22 and a northward gradient of 0.31; vibration characteristics include a vibration coupling index of 0.48, a normalized spectral energy of 0.29 in the lower vibration frequency band, and a normalized spectral energy of 0.17 in the higher vibration frequency band. Based on this, the actual line-of-sight error for the current cycle is calculated to be 1.93°, and the scanning direction mismatch is 4.28°.

[0155] Within the 100 ms prediction window, the preceding two steps correspond to significant changes in the dam slope normal. Therefore, the prediction controller outputs the first compensation command as a roll axis of 0.118 rad / s and a pitch axis of -0.146 rad / s. After superimposing the fast feedback, the final execution command is a roll axis of 0.129 rad / s and a pitch axis of -0.159 rad / s. After execution, the line-of-sight error in the next cycle is reduced to 1.21°, and the scan direction mismatch is reduced to 2.96°, stabilizing below 0.89° and 2.1° respectively in the three cycles after the turn. This process demonstrates that, as the dam slope bend approaches, the changes in the terrain normal and the scan direction mismatch are already reflected in the control commands, thereby reducing strip stitching errors and inconsistencies in ground projection speed during the acquisition phase.

[0156] This exemplary simulation uses a digital scenario of a tailings dam slope with bends, with a simulation duration of 480 seconds. The normal operating condition is a 2 m / s crosswind, while the impact condition includes a 0.3 s lateral gust in the turning area and occasional local distance measurement packet loss. The control scheme uses fixed-attitude flight with post-flight geometric correction. The simulation was repeated 8 times, and the average results are as follows: the control scheme has an average strip splicing overlap error of 1.26 pixels, an effective spectral pixel ratio of 91.3%, an average strip orthogonality deviation of 5.8°, and a standard deviation of inter-row overlap deviation of 0.87 pixels; the present invention's scheme has an average strip splicing overlap error of 0.58 pixels, an effective spectral pixel ratio of 96.8%, an average strip orthogonality deviation of 2.4°, and a standard deviation of inter-row overlap deviation of 0.39 pixels. Eight flight tests were completed in a tailings dam slope area, with a total flight time of 52 minutes and 22 effective bend segments. The control procedure employed alternating flights on the same platform, with four flights utilizing the attitude stabilization control of this invention and four flights using fixed attitude flight with post-flight geometric correction. Statistical results are as follows: the control scheme had a mean stripe splicing overlap error of 1.18±0.21 pixels, an effective spectral pixel percentage of 92.1%±1.4%, a mean stripe orthogonality deviation of 4.96°±0.63°, and a mean ground projection velocity consistency index of 0.14±0.03; the invention scheme had a mean stripe splicing overlap error of 0.64±0.12 pixels, an effective spectral pixel percentage of 96.2%±0.9%, a mean stripe orthogonality deviation of 2.71°±0.44°, and a mean ground projection velocity consistency index of 0.07±0.02.

[0157] Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Embodiment 1 to complete operations such as abnormal degradation processing, state update and offline parameter version verification. Its specific execution logic is the same as that in Embodiment 1, and will not be repeated here.

[0158] Example 4: A Hyperspectral Mapping Attitude Stabilization Control System for Complex Topography in Mining Areas

[0159] This embodiment provides a hyperspectral mapping attitude stabilization control system for complex terrain in mining areas to implement the aforementioned method, including a multi-source acquisition module, a terrain normal construction module, a line-of-sight error generation module, a prediction compensation module, an anomaly degradation and quality management module, and a closed-loop parameter update module.

[0160] The multi-source acquisition module is used to collect body attitude data, body angular velocity data, gimbal attitude data, terrain elevation data, ranging data, and load vibration data. It also uses the exposure time of the hyperspectral camera as a unified time axis to align the collected multi-source data with time stamps and outputs unified time-stamped data frames.

[0161] The terrain normal construction module is used to construct a local surface elevation window based on the aligned terrain elevation data and ranging data. The terrain gradient is calculated from the local surface elevation window and a local surface normal vector is constructed. The ideal load line-of-sight reference is determined based on the local surface normal vector.

[0162] The line-of-sight error generation module is used to solve the actual load line-of-sight based on the body attitude matrix, gimbal attitude matrix and installation extrinsic parameter matrix, calculate the line-of-sight error angle and scanning direction mismatch, and construct an enhanced state vector containing vibration characteristics.

[0163] The prediction compensation module is used to solve the gimbal compensation control sequence in a rolling manner within a preset prediction time domain based on a discrete prediction model that includes the current error state term, terrain gradient change term, and vibration disturbance feature term, and to generate gimbal compensation instructions. The prediction compensation module can also switch to a simplified solution mode when the processor load rate exceeds a preset threshold or when a solution timeout trend occurs. This function is a preferred engineering implementation method.

[0164] The anomaly degradation and quality management module is used to switch to safe degradation control when the normal reference fails, the prediction solution is abnormal, the encoder is abnormal, or the continuous prediction error exceeds the preset conditions, and outputs quality mark information when the ranging data and terrain elevation data fail at the same time; the quality mark information can be associated with the exposure row number index for subsequent strip stitching and data filtering.

[0165] The closed-loop parameter update module is used to read new line-of-sight error, gimbal response status and exposure synchronization information after the current control cycle is executed to update the enhanced state vector, cache state transition data, and perform offline parameter updates and version rollback management according to the task log. Among them, the parameter update involves at least terrain change propagation parameters, vibration propagation parameters and feedback gain parameters, and the version rollback is determined based on the average line-of-sight error, the high quantile statistic of line-of-sight error and the mean value of scan direction mismatch on the independent verification log.

[0166] The modules described above are connected sequentially according to the data flow: multi-source acquisition, terrain normal construction, actual line-of-sight calculation, error state generation, prediction compensation, anomaly degradation, and closed-loop parameter update. Each module can be implemented programmatically by the same airborne processor, or it can be implemented in a distributed collaborative manner by the flight control processor and the payload processor, but the data flow and control logic remain consistent. Through this system architecture, terrain reference construction, line-of-sight attitude stabilization control, anomaly handling, and parameter management in complex terrain mapping can be integrated into a unified control link.

[0167] Core Symbolic Variable Description Table

[0168] The core symbolic variables involved in this invention are described in detail in the table below.

[0169]

[0170]

[0171] It should be noted that relational terms such as "first" and "second" are used only to distinguish different entities or steps and do not indicate an actual order or hierarchy; terms such as "including" and "comprising" are open-ended and do not exclude other elements not explicitly listed. The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, any modifications, equivalent substitutions, and improvements made without departing from the principles and spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for attitude stabilization control in hyperspectral mapping of complex terrain in mining areas, applied to a flight platform equipped with a hyperspectral camera and a servo gimbal, performing load line-of-sight attitude stabilization adjustment based on multi-source sensor data, characterized in that... include: The system collects data on the flight platform's attitude, angular velocity, gimbal attitude, terrain elevation, ranging, and load vibration. It also aligns the collected multi-source data with the exposure time of the hyperspectral camera as a unified time axis. A local surface elevation window is constructed based on the aligned terrain elevation data and distance measurement data. The terrain gradient is calculated and a local surface normal vector is constructed based on the local surface elevation window. The ideal load line-of-sight reference is determined based on the local surface normal vector. Simultaneously, the actual load line of sight is calculated based on the airframe attitude matrix, gimbal attitude matrix, and installation extrinsic parameter matrix. The line of sight error angle between the actual load line of sight and the ideal load line of sight reference is calculated, and the scanning direction mismatch is calculated in combination with the ground speed direction of the flight platform and the local contour line tangent. An enhanced state vector is constructed based on at least attitude information, error state information, terrain information, relative ground height information, and vibration feature information extracted from load vibration data. A discrete prediction model is then established that includes the current error state term, terrain gradient change term, and vibration disturbance feature term. Within the preset prediction time domain, the gimbal compensation control sequence is solved in a rolling manner according to the discrete prediction model. The optimization objective of the gimbal compensation control sequence simultaneously constrains the line-of-sight error, the rate of change of line-of-sight error, the scanning direction mismatch, the control energy, the control smoothness, and the mechanical boundary risk. The optimal control initial value of the current cycle is superimposed with the fast feedback compensation value to generate the gimbal compensation command. The gimbal compensation command is sent to the gimbal to perform attitude compensation, and the system switches to safety degradation control when the normal reference fails, the prediction solution is abnormal, the encoder is abnormal, or the continuous prediction error exceeds the preset mismatch condition. After executing the current control cycle, read the new line-of-sight error, gimbal response status and exposure synchronization information, update the enhanced state vector for the next cycle, and cache the state transition data of the most recent control cycle for offline parameter recalibration after the flight ends; The offline parameter recalibration is used to update the terrain change propagation parameters, vibration propagation parameters, and feedback gain parameters. The parameter versions before and after the update are compared through independent verification logs. If the preset improvement conditions are not met, the parameters are rolled back.

2. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 1, characterized in that, The step of aligning the acquired multi-source data with the exposure time of the hyperspectral camera as a unified time axis includes: Map flight control attitude data and gimbal attitude data to exposure timestamps; Extract spectral features from load vibration data within a preset control period; Perform local window clipping on the digital elevation model based on the current location projection; When ranging data fails, either retain the most recent valid value or write an invalid flag based on the duration of the failure.

3. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 1, characterized in that, The step of constructing a local surface elevation window based on aligned terrain elevation data and distance measurement data, and calculating the terrain gradient and constructing a local surface normal vector from the local surface elevation window includes: Based on the validity of the ranging data and its deviation relationship with the slope distance calculated by the digital elevation model, the local elevation construction method is determined among ranging correction of the digital elevation model, prioritizing the use of the digital elevation model, using ranging to back-calculate the center elevation of the window, and weighted fusion of the two, and outlier removal is performed on abnormal ranging values. Based on the local surface elevation window, differential calculations are performed in two horizontal orthogonal directions to obtain the terrain gradient; Based on the terrain gradient, an unnormalized normal vector is constructed and normalized to obtain a local surface normal vector; When the elevation window is located at a missing boundary, resulting in insufficient effective grid, the terrain gradient is obtained by using one-sided difference or the nearest effective value.

4. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 3, characterized in that, The process of determining the ideal load line-of-sight reference based on the local surface normal vector, and calculating the actual load line-of-sight based on the aircraft attitude matrix, gimbal attitude matrix, and installation extrinsic parameter matrix, includes: In near-normal mapping mode, the reverse direction of the local surface normal vector is determined as the ideal load line-of-sight reference; In a mapping mode that takes forward safety into account, the ideal load line of sight reference is determined based on the local surface normal vector reversal, the track tangential direction, and the forward safety distance. Based on the product relationship between the body attitude matrix, the gimbal attitude matrix, and the installation extrinsic parameter matrix, the nominal optical axis of the load is transformed into the navigation coordinate system to obtain the actual load line of sight. The line-of-sight error angle is determined based on the angle between the actual load line-of-sight and the ideal load line-of-sight reference, and the scanning direction mismatch is determined based on the angle between the horizontal projection direction of the flight platform ground speed and the tangent of the local contour line. When the terrain slope is less than the preset slope threshold, reduce the weight of the scanning direction mismatch in the control or set it to a low-impact state.

5. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 1, characterized in that, The vibration features extracted from the load vibration data include at least the vibration coupling index and the frequency band normalized spectral energy, wherein: The vibration coupling index is calculated from the normalized correlation strength between the vibration acceleration spectrum of the load base and the angular velocity spectrum of the body; The normalized spectral energy of the frequency band includes the normalized spectral energy of the lower vibrational frequency band and the normalized spectral energy of the higher vibrational frequency band. The rate of change of line-of-sight error is estimated by the difference between the line-of-sight error angle of the current period and the smoothed line-of-sight error angle of the previous period; The enhanced state vector also includes state markers used to characterize the validity of spectral samples or to characterize vibrational features following the previous cycle.

6. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 1, characterized in that, The establishment of a discrete prediction model that includes the current error state term, terrain gradient change term, and vibration disturbance characteristic term includes: The predicted state is defined as an error state vector consisting of the line-of-sight error angle, the line-of-sight error rate of change, and the scan direction mismatch. The state evolution matrix is ​​used to characterize the natural evolution of the current error state to the error state in the next cycle; The influence of gimbal roll compensation command and gimbal pitch compensation command on the error state of the next cycle is characterized by a control input matrix. The topographic change propagation matrix is ​​used to characterize the propagation effect of future topographic gradient changes on the error state of the next cycle, and the vibration propagation matrix is ​​used to characterize the propagation effect of vibration disturbance characteristics on the error state of the next cycle.

7. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 6, characterized in that, The step of rolling the solution of the gimbal compensation control sequence within the preset prediction time domain includes: Based on the current position, aircraft heading, ground speed vector and local terrain data, the terrain gradient of future path points within the prediction window is extrapolated to form a future terrain gradient sequence. A rolling optimization objective function is constructed using the following terms: line-of-sight error, line-of-sight error rate of change, scan direction mismatch, control energy, control smoothing, and risk. The rolling optimization objective function is solved using a constrained solver to obtain the gimbal compensation control sequence corresponding to multiple future control cycles; Only the current optimal control initial value in the gimbal compensation control sequence is executed, and the prediction and solution are re-executed in the next control cycle; When the processor load rate exceeds the preset load rate threshold or a timeout trend occurs, switch to simplified solution mode.

8. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 7, characterized in that, The risk term includes at least one of the following characterization constraints: The risk that the future line-of-sight error will exceed the preset safety error threshold; the risk that the future scan direction mismatch will exceed the preset mismatch safety threshold; The risk that the future gimbal's pitch angle will approach its mechanical limit; The risk that the roll angle of the gimbal will approach its mechanical limit in the future; In the optimization calculation, the angle, angular velocity, and angular position quantities are all uniformly expressed in radians.

9. The attitude stabilization control method for hyperspectral mapping of complex terrain in mining areas according to claim 1, characterized in that, The step of superimposing the optimal control initial value of the current cycle with the fast feedback compensation value to generate a gimbal compensation command, and switching to safe degradation control in case of an anomaly, includes: The fast feedback compensation amount is calculated based on the current line-of-sight error angle and the line-of-sight error change rate, and then the fast feedback compensation amount is superimposed with the optimal control initial value of the current cycle to obtain the final gimbal compensation command. An angular velocity limit is applied to the final gimbal compensation command, and the mechanical limit of the gimbal angular position is checked in combination with the angular velocity integral result; When the normal reference is continuously invalid, the prediction solution times out continuously, the gimbal encoder is abnormal, or the continuous prediction line-of-sight error exceeds the preset error threshold, the terrain prediction term and vibration propagation term are turned off, and only conservative feedback control based on the current error is retained. When both ranging data and terrain elevation data fail, the ideal load line reference is switched to the direction directly below the aircraft, and quality markings are applied to the corresponding exposure lines.

10. A hyperspectral mapping attitude stabilization control system for complex terrain in mining areas, characterized in that, include: The multi-source acquisition module is used to collect body attitude data, body angular velocity data, gimbal attitude data, terrain elevation data, ranging data, and load vibration data, and to align the acquired multi-source data with the exposure time of the hyperspectral camera as a unified time axis. The terrain normal construction module is used to construct a local surface elevation window based on the aligned terrain elevation data and ranging data, calculate the terrain gradient and construct a local surface normal vector based on the local surface elevation window, and determine the ideal load line-of-sight reference based on the local surface normal vector. The line-of-sight error generation module is used to solve the actual load line-of-sight based on the body attitude matrix, gimbal attitude matrix and installation extrinsic parameter matrix, calculate the line-of-sight error angle and scanning direction mismatch, and construct an enhanced state vector containing vibration characteristics. The prediction and compensation module is used to solve the gimbal compensation control sequence in a rolling manner within a preset prediction time domain based on a discrete prediction model that includes the current error state term, terrain gradient change term, and vibration disturbance characteristic term, and to generate gimbal compensation instructions. The abnormal degradation and quality management module is used to switch to safe degradation control when the normal reference fails, the prediction solution is abnormal, the encoder is abnormal, or the continuous prediction error exceeds the preset mismatch condition, and outputs quality mark information when the ranging data and terrain elevation data fail at the same time. The closed-loop parameter update module is used to read new line-of-sight error, gimbal response status and exposure synchronization information after the current control cycle is executed to update the enhanced state vector, cache state transition data, and perform offline parameter updates and version rollback management based on the task log.

Images