An imaging quality constrained flight control system

Abstract

The embodiment of the disclosure provides an imaging quality constrained flight control system, in which: a quality evaluation module is used for generating an imaging quality vector containing signal-to-noise ratio, phase noise, strip overlap degree and point cloud sparsity based on millimeter wave echo and point cloud data; a constraint mapping module is used for generating cost weight, soft constraint penalty coefficient, hard constraint threshold and safety boundary tightening coefficient through a mapping function based on the imaging quality vector; a rolling optimization module is used for constructing a quality-aware cost function based on an aircraft dynamics model and the output of the constraint mapping module, and solving an optimal control sequence in a limited time domain combined with geometric consistency and coverage constraint; and an execution interface module is used for converting the first control quantity in the optimal control sequence into an executable driving command and issuing the executable driving command to a UAV propulsion system and a gimbal. The system can achieve the effects of imaging stability improvement, coverage integrity improvement, robustness improvement and efficiency improvement.

Description

Technical Field

[0001] This disclosure relates to the fields of unmanned aerial vehicle (UAV) flight control, millimeter-wave radar imaging, and UAV path planning and optimization, and in particular to an imaging quality-constrained flight control system. Background Technology

[0002] Online inspection of confined spaces such as industrial boilers requires obtaining stable imaging results of heated surfaces in environments with high temperatures, strong smoke and dust, strong convection, and strong reflection. Millimeter-wave radar has advantages in such scenarios, including penetrating smoke and dust, insensitivity to contrast, and adaptability in size and power consumption, and has gradually replaced pure vision solutions as the core sensor. UAVs, as mobile carriers, provide flexible near-wall coverage capabilities and highly maneuverable trajectory execution capabilities. However, in areas such as water-cooled walls, horizontal flues, and tail flues, the narrow space, complex wall curvature, dense array of steel components, and strong updrafts and backflow vortices cause the flight control and imaging processes to be coupled: attitude disturbances change the antenna incident angle, velocity fluctuations change the synthetic aperture sampling density, the flight path spacing determines strip overlap and stitching errors, and the wall distance and normal deflection angle jointly affect the echo intensity and phase stability.

[0003] The commonality of existing solutions lies in the "decoupling of perception and control." A common approach is to first use a coverage planning algorithm to provide a regular grid or serpentine path, and then implement tracking control using a cascaded proportional-integral-derivative controller (PID controller) or a linear quadratic regulator (LQR). Imaging quality is then controlled by offline thresholding or post-processing quality checks. To mitigate environmental disturbances, some studies have added fusion localization between an inertial measurement unit (IMU) and a distance sensor to suppress attitude and wall distance fluctuations. To reduce the risk of missed detections, some solutions uniformly increase overlap and reduce redundancy at the path level, at the cost of decreased efficiency and endurance. Other active sensing methods attempt to drive local revisits using information entropy or feature uncertainty, but these are mostly based on cameras or LiDAR, making it difficult to directly measure key quality factors in millimeter-wave imaging.

[0004] The weakness of this decoupling paradigm lies in its inability to drive control decisions using "imaging quality." First, quality metrics are lacking. The signal-to-noise ratio, phase noise, stripe overlap, and point cloud sparsity of millimeter-wave imaging have clear physical and geometric meanings, directly determining target separability and stitching stability. However, current control targets mostly revolve around trajectory errors and energy costs, failing to quantify quality as a differentiable or constrainable control variable. Second, the trade-off between velocity and resolution is not internalized. The azimuth resolution of synthetic aperture radar varies with the equivalent aperture and sampling density; velocity and scan rate directly determine the sampling rhythm. Current control does not explicitly penalize the "too fast, leading to sparsity" imaging degradation at the optimization level. Third, surface geometric consistency constraints are insufficient. Boiler water-cooled walls are periodic tube-pile curved surfaces; small deviations in the normal and wall distance can be amplified into systematic errors in the incident angle and echo phase. While path-level coverage can be increased, if geometric consistency constraints are not applied at the control level, local "unclear" or "unstable" imaging holes will still appear. Fourth, the disturbance loop is not closed. The random forces and torques caused by the strong rising smoke cause related drifts in attitude, speed and gimbal pointing. If the quality degradation cannot be penalized or the safety boundary tightened in real time at the optimization layer, cascading degradation of "the more you shoot, the blurrier it gets" is likely to occur. Summary of the Invention

[0005] This disclosure provides an imaging quality-constrained flight control system to address the problems in existing flight control schemes, such as the lack of quality measurement, the lack of endogenization of the speed and resolution trade-off, insufficient surface geometric consistency constraints, and non-closed disturbance loops caused by the decoupling of perception and control.

[0006] In view of the above problems, a first aspect is to provide a flight control system with imaging quality constraints, the system comprising: a quality assessment module, a constraint mapping module, a rolling optimization module, and an execution interface module;

[0007] The quality assessment module is used to generate an imaging quality vector that includes signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity based on millimeter-wave echo and point cloud data.

[0008] The constraint mapping module is used to generate cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients based on the imaging quality vector and through a mapping function.

[0009] The rolling optimization module is used to construct a quality-aware cost function based on the aircraft dynamics model, the cost weight, the soft constraint penalty coefficient, the hard constraint threshold and the safety boundary tightening coefficient, and to solve for the optimal control sequence in the finite time domain by combining geometric consistency and coverage integrity constraints.

[0010] The cost weight is used to adjust the importance ratio of the trajectory tracking term and the control smoothing term in the quality-perceived cost function; the soft constraint penalty coefficient is used to control the penalty intensity of the obstacle term in the quality-perceived cost function; the hard constraint threshold is used to limit the feasible boundary of the key parameters of UAV flight, which include at least one of velocity, angular velocity, acceleration, wall distance, point cloud density, and strip step size; the safety boundary tightening coefficient is used to dynamically reduce the feasible domain of the hard constraint threshold when the imaging quality deteriorates.

[0011] The execution interface module is used to convert the first control quantity in the optimal control sequence into an executable drive command and send it to the UAV propulsion system and gimbal.

[0012] In conjunction with the first aspect, in one possible implementation, the quality assessment module is used for:

[0013] The signal-to-noise ratio is calculated by the ratio of the power of the main lobe sample set after the distance-directed FFT to the power of the noise window set.

[0014] For the mean-free phase sequence of complex echoes of K consecutive pulses within the same line of sight, the standard deviation is calculated after unwrapping, and this standard deviation is determined as phase noise.

[0015] The effective strip width is determined based on the wall distance, incident angle, and half-power beamwidth; the strip overlap is calculated based on the effective strip width and the lateral distance from the centerline of the current strip to the centerline of the previous strip.

[0016] The point cloud is counted within a unit surface arc length, and the statistical point cloud density is obtained by normalizing it according to the grid area. The point cloud sparsity is determined based on the ratio of the statistical point cloud density to the reference density.

[0017] The signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity are normalized. The normalized signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity are used as components of the imaging quality vector, and the imaging quality vector is output.

[0018] In conjunction with the first aspect, in one possible implementation, the quality assessment module is used for:

[0019] According to the family of monotonically differentiable functions ;

[0020] The original mass vector Mapping to the [0,1] interval yields the normalized image quality vector. ;

[0021] in, For signal-to-noise ratio, For phase noise, For strip overlap, Point cloud sparsity; Original mass vector The first in i One portion, Normalized image quality vector The i-th component in Original mass vector The i The normalization function for each component; This is the adjustment coefficient of the signal-to-noise ratio normalization function. This serves as the baseline value for the signal-to-noise ratio. This is the adjustment coefficient for the phase noise normalization function. This is the reference value for phase noise.

[0022] In conjunction with the first aspect, in one possible implementation, the constraint mapping module is used for:

[0023] According to the cost weight mapping formula Determine the cost weights ;

[0024] in, For the imaging quality vector, For benchmark weights, Here is the gain matrix. As the lower bound of the cost weight, The upper limit of the cost weight; cost weight The components include at least one of the following: position tracking weight, velocity weight, acceleration weight, angular velocity weight, angular acceleration weight, lateral deviation weight, wall distance deviation weight, phase noise exceeding limit weight, and point cloud density insufficient weight.

[0025] In conjunction with the first aspect, in one possible implementation, the cost weight mapping formula satisfies at least one of the following rules:

[0026] Imaging quality vector The sparsity component of point cloud When the cost weight increases, velocity weights in With acceleration weight improve;

[0027] Imaging quality vector Phase noise components When the cost weight increases, Angular velocity weights in With angular acceleration weight improve;

[0028] Imaging quality vector The strip overlap component When reduced, cost weight Lateral deviation weight in improve;

[0029] Imaging quality vector Signal-to-noise ratio components in When reduced, cost weight Wall distance deviation weight improve.

[0030] In conjunction with the first aspect, in one possible implementation, the constraint mapping module is used for:

[0031] Using a linear tightening formula Determine the hard constraint threshold ;

[0032] in, K is the initial hard constraint threshold, and K is the safety boundary tightening coefficient;

[0033] This is a quality degradation characterization vector. , Image quality vector The signal-to-noise ratio component in Image quality vector Phase noise components in Image quality vector The strip overlap component in the data. Image quality vector The point cloud sparsity component;

[0034] Hard constraint threshold The components include the upper speed limit. Upper limit of angular velocity , acceleration upper limit Minimum wall distance Maximum wall distance Minimum effective point cloud density Maximum step distance At least one of them;

[0035] The linear tightening formula satisfies the rule that when the imaging quality deteriorates, the feasible boundary of the key flight parameters of the UAV shrinks.

[0036] In conjunction with the first aspect, in one possible implementation, the constraint mapping module is used for:

[0037] when When the value is less than the first segment threshold, the safety boundary tightening coefficient K adopts the first value;

[0038] when When the threshold of the first segment is exceeded but the threshold of the second segment is not exceeded, the quality degradation mode is triggered, and the safety boundary tightening coefficient K adopts the second value.

[0039] when When the second segment threshold is exceeded, the disturbance over-limit mode is triggered, and the safety boundary tightening coefficient K adopts the third value;

[0040] The second value is greater than the first value, and the third value is greater than the second value.

[0041] In conjunction with the first aspect, in one possible implementation, the constraint mapping module is used for:

[0042] According to the soft constraint penalty mapping formula Determine the soft constraint penalty coefficient ;

[0043] in, Here, H represents the initial soft constraint penalty coefficient, and H is the adjustment matrix. This is a quality degradation characterization vector;

[0044] Soft constraint penalty coefficient The components include at least one of the following: wall distance deviation from obstacle weight, lateral deviation weight, insufficient point cloud density weight, and phase noise exceeding limit weight;

[0045] The soft constraint penalty mapping formula satisfies the rule that when the imaging quality deteriorates, the constraint stiffness of the soft constraint penalty is increased.

[0046] In conjunction with the first aspect, in one possible implementation, the rolling optimization module is used to:

[0047] Adopting a linearizable aircraft dynamics model Construct a quality-aware cost function

[0048] ;

[0049] The prediction domain is the discrete time domain, and the step size is... The prediction domain length is N ;

[0050] For the first k The state at each time step , For the first k The location of the drone at each time step. For the first k The drone speed at each time step For the first k Attitude parameters or small-angle attitude vectors at each time step;

[0051] For the first k Control quantity per time step , For the first k The expected body acceleration at each time step For the first k The gimbal angular velocity at each time step;

[0052] For the first k The reference trajectory position at each time step For location tracking weights;

[0053] For the first k Reference trajectory velocity at each time step Speed ​​weights;

[0054] For the first k The angular velocity of the drone at each time step As the weight for angular velocity, For acceleration weights;

[0055] For the first k The distance between the drone and the wall of the measured curved surface at each time step. For the desired observation distance, As the weight for wall distance deviation, The smooth hinge penalty function for wall distance deviation;

[0056] For the first k Local lateral step size of adjacent stripes at each time step. For the maximum step size, , The smooth hinge penalty function for lateral deviation;

[0057] For the first k Effective point cloud density at each time step To achieve the minimum effective point cloud density, The point cloud density is insufficient for weighting. For insufficient point cloud density, a smooth hinge penalty function is used.

[0058] For the first k Phase noise proxy at each time step The maximum phase noise threshold. For phase noise exceeding the limit weight, The smooth hinge penalty function is used to compensate for excessive phase noise.

[0059] Position tracking weights Speed ​​weight Angular velocity weight Acceleration weight The cost weight configuration output by the constraint mapping module;

[0060] Wall distance deviation weight Horizontal deviation weight Insufficient point cloud density weight Phase noise exceeding limits weight The penalty coefficient for soft constraints is determined by the output of the constraint mapping module;

[0061] Maximum step distance Maximum phase noise threshold Minimum effective point cloud density The hard constraint threshold and safety boundary tightening coefficient output by the constraint mapping module are used to determine the safety boundary.

[0062] In conjunction with the first aspect, in one possible implementation, the rolling optimization module is used to:

[0063] The following constraints are added as geometric consistency and coverage integrity constraints:

[0064] ;

[0065] Wherein, the normal of the measured surface is The sensor's line of sight is , The minimum stripe overlap threshold. For effective strip width, Angle of incidence;

[0066] Minimum wall distance Maximum wall distance Maximum incident angle threshold Minimum strip overlap threshold Maximum step distance Minimum effective point cloud density Speed ​​limit Upper limit of angular velocity , acceleration upper limit The hard constraint threshold and safety boundary tightening coefficient output by the constraint mapping module are updated in real time.

[0067] The beneficial effects of the embodiments disclosed herein include:

[0068] Quality endogenization: Normalize SNR, phase noise, strip overlap, and point cloud sparsity into differentiable quality vectors, and simultaneously map them to cost, soft constraints, hard constraints, and safety boundaries to achieve isomorphic representation of quality and control.

[0069] Geometric consistency control: Using the wall distance and the angle between the normals as the core constraints, the surface geometry is introduced into the feasible region to maintain the stability of the incident geometry and reduce phase drift and splicing distortion.

[0070] Coverage and sampling coordination: Using parsing proxies of strip step size and point cloud density, overlap and sampling density are incorporated into optimization objectives and constraints, avoiding the inefficient practice of "compensating for quality by encrypting paths".

[0071] Boundary scheduling disturbance resistance: When the quality deteriorates or the disturbance exceeds the threshold, the upper limits of speed, attitude and distance are automatically tightened to form a quality-priority safety boundary scheduling to suppress cascading degradation.

[0072] In boiler inspection under deep peak-shaving conditions, this disclosure offers the following advantages: improved imaging stability with a significant reduction in blurred frames and stitching seams in low SNR and high phase noise segments; improved coverage integrity with reduced strip misses and holes; increased efficiency with reduced revisits and redundant flights at the same quality threshold; and improved robustness, achieving dual-loop convergence for both quality and safety in the face of attitude disturbances caused by updrafts and backflow vortices. The system is modularly implemented, facilitating integration with existing UAVs and millimeter-wave payloads, and can be extended to other confined, low-light environment inspection tasks. Attached Figure Description

[0073] Figure 1 A schematic diagram of the structure of a flight control system with imaging quality constraints provided in an embodiment of this disclosure;

[0074] Figure 2 An imaging quality assessment flowchart provided for embodiments of this disclosure;

[0075] Figure 3 A mapping diagram of imaging quality to constraints and cost weights provided for embodiments of this disclosure;

[0076] Figure 4 This is a flowchart of flight control and rolling execution provided for an embodiment of the present disclosure. Detailed Implementation

[0077] This disclosure provides an imaging quality-constrained flight control system. Preferred embodiments of this disclosure are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of this disclosure. Furthermore, the embodiments and features described herein can be combined with each other unless otherwise specified.

[0078] In existing flight control schemes, the bottleneck lies in the bidirectional coupling of sampling quality and control law, rather than a one-way subordination. The imaging system needs to achieve sufficient density and consistency in sampling both spatially and temporally; the control system needs to provide trajectories and attitudes that meet the geometric and statistical requirements of sampling under dynamic and safety constraints. The lack of a unified target carrier between the two prevents the optimizer from automatically adjusting the cost structure and feasible region when quality deteriorates, making it difficult to simultaneously achieve the four objectives of stability, clarity, continuity, and efficiency.

[0079] The industry trend has emerged towards mission-oriented flight control and proactive sampling based on quality perception. On one hand, Model Predictive Control (MPC) is suitable for handling concurrent problems involving nonlinearity, saturation, collisions, and both hard and soft constraints within a finite time domain. It can transform information or quality objectives into cost and constraint terms, enabling interpretable online trade-offs. On the other hand, the advantages of millimeter waves in smoke and dust environments require a matching set of quality indicators and mapping mechanisms, allowing the controller to understand the "good and bad" aspects of millimeter waves, rather than reusing heuristic indicators from vision or lasers. Edge-to-edge collaboration, equipped with computing power and real-time optimization algorithms, also provides the computational power and architectural conditions for a quality closed loop.

[0080] Against this backdrop, this invention proposes a flight control method with imaging quality constraints. This method uses the imaging quality vector generated from millimeter-wave echoes and point clouds as primary control variables. By mapping quality to constraints and cost weights, quality requirements are injected into the optimization problem. A finite-time domain optimization model incorporating attitude, velocity, acceleration, and scanning speed is constructed. Combined with wall distance measurement and surface geometric consistency, executable soft and hard constraints are formed, and online solution and execution are performed using rolling optimization. The goal is to achieve a synchronous closed loop of quality and safety within a confined space, enabling the UAV to complete high-quality coverage imaging with minimal redundancy, providing a stable and reliable data foundation for boiler heating surface defect detection.

[0081] This invention belongs to the interdisciplinary field of deep integration of flight control and millimeter-wave radar imaging. In this invention, millimeter-wave radar imaging provides a reliable detection and imaging means in environments with interference such as smoke and dust; the control algorithm excels at handling multiple constraints during flight, achieving optimized real-time trajectory control; and UAV path planning and optimization technology ensures that the aircraft can safely and efficiently complete detection tasks in confined spaces. Specifically, this invention provides a flight control method with imaging quality constraints to improve the autonomous flight stability and imaging quality of UAVs in complex and confined spaces. It particularly relates to the application of this method in intelligent detection tasks in typical enclosed and confined spaces such as industrial boilers. This method introduces millimeter-wave radar imaging quality feedback to optimize the control of the UAV's flight trajectory and attitude, overcoming flight control challenges such as smoke and dust interference, complex wall structures, and limited space. This invention combines flight control with imaging feedback to enhance the autonomous flight capability and detection imaging effect of UAVs in special industrial environments.

[0082] This invention addresses the key challenges of UAV millimeter-wave detection missions within confined industrial spaces (such as boiler furnaces, water-cooled walls, and flues) by solving the following problems: First, under conditions of strong disturbances and spatial constraints, how to transform imaging quality from a "post-evaluation indicator" into a real-time control objective and constraint, enabling flight trajectory, attitude, and sensor scanning behavior to react instantly to quality degradation. Second, how to simultaneously satisfy dynamic feasibility, geometric safety, coverage integrity, and sampling sufficiency within a finite time domain, while avoiding efficiency losses due to redundant revisits. Third, how to achieve coordinated convergence of quality and safety under conditions of flue gas disturbances and echo fluctuations through interpretable weighting and boundary scheduling.

[0083] Figure 1 This is a schematic diagram of the structure of a flight control system with imaging quality constraints provided in an embodiment of this disclosure, as shown below. Figure 1 As shown, flight control is driven by an imaging quality closed loop: The UAV is equipped with a millimeter-wave radar and gimbal 201, an IMU (Inertial Measurement Unit) and a wall distance sensor 202. An imaging quality-constrained flight control system 100 (including a quality assessment module 101, a constraint mapping module 102, a rolling optimization module 103, and an execution interface module 104) runs on an edge computing unit. In each control cycle, the system collects echo and inertial navigation data, calculates the imaging quality vector, maps the quality vector to cost weights, soft and hard constraints, and safety boundaries, solves for the optimal control quantity within a finite time domain, and sends it to the UAV propulsion system 301 and the gimbal. This process is repeated in the next cycle.

[0084] Specifically, Figure 1The arrows in the diagram indicate the directions of the sensing data stream and the control command stream. The sensor data stream enters the quality assessment module 101 and outputs an imaging quality vector; the imaging quality vector enters the constraint mapping module 102 to generate cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients; the rolling optimization module 103 solves for the optimal control sequence based on aircraft dynamics, the output of the constraint mapping module 102, and geometric constraints; the execution interface module 104 distributes the first control quantity in the optimal control sequence to the UAV propulsion system 301 and the gimbal in the millimeter-wave radar and gimbal 201, while recording the execution status for use in the next cycle.

[0085] The following is a detailed description of each module in a flight control system 100 with imaging quality constraints.

[0086] The quality assessment module 101 is used to generate an imaging quality vector that includes signal-to-noise ratio, phase noise, strip overlap and point cloud sparsity based on millimeter-wave echo and point cloud data.

[0087] In this embodiment of the disclosure, in each control cycle Based on millimeter-wave echo and point cloud generation, the signal-to-noise ratio is estimated. Phase noise Strip overlap With point cloud sparsity Four types of quality elements, denoted as the original quality vector. .

[0088] Define a monotonically differentiable normalized mapping Scale each component to And maintain physical monotonicity: ;

[0089] Among them, the signal-to-noise ratio component A larger value indicates better signal-to-noise ratio and phase noise. A larger value indicates poorer phase stability and a higher stripe overlap component. Indicates the sufficiency of stripe overlap, and the sparsity components of the point cloud. This indicates the sparsity level of the point cloud.

[0090] Figure 2 An imaging quality assessment flowchart is provided for embodiments of this disclosure, such as... Figure 2 As shown, the millimeter-wave echo data, IMU and wall distance data are input, and after preprocessing (synchronization and buffering), spectrum analysis and point cloud generation, four types of indicators are estimated: estimated signal-to-noise ratio, phase noise, coverage area strip overlap, and point cloud sparsity. These are combined into an original quality vector, which is then normalized to obtain an imaging quality vector and output to the constraint mapping module 102.

[0091] The estimation of four types of indicators—signal-to-noise ratio, phase noise, overlap of covered area stripes, and sparsity of point cloud—is explained in detail below.

[0092] As one possible implementation, the quality assessment module 101 is used to estimate the signal-to-noise ratio, phase noise, stripe overlap, and point cloud density in the following ways.

[0093] 1. Signal-to-noise ratio estimation: The signal-to-noise ratio is calculated by the ratio of the power of the main lobe sample set after the distance-directed FFT to the power of the noise window set.

[0094] Specifically, let the main lobe sampling set after the distance-to-FFT be denoted as . The noise window set is , No. r The amplitude of each sampling point is .

[0095] The signal power is determined according to the following formula. and noise power :

[0096] ;

[0097] The signal-to-noise ratio (SNR) is calculated using the following formula:

[0098] .

[0099] The main lobe region is determined by the target threshold and gimbal pointing of the previous cycle, and the noise window excludes strong scattering range cells.

[0100] 2. Phase noise estimation: For the mean-free phase sequence of complex echoes of K consecutive pulses within the same line of sight, the standard deviation is calculated after unwrapping, and this standard deviation is determined as the phase noise.

[0101] Specifically, for continuous lines of sight within the same line of sight The complex echo of each pulse is taken as the reference range unit. Define the mean-removed phase sequence , To obtain the mean phase sequence The mean is calculated using the following formula, and the standard deviation is then determined. This standard deviation is defined as the phase noise. :

[0102] ;

[0103] Phase noise A larger value indicates phase instability. Reference distance unit. Adaptive selection based on wall distance and echo intensity.

[0104] 3. Strip overlap estimation: The effective strip width is determined based on the wall distance, incident angle, and half-power beamwidth; the strip overlap is calculated based on the effective strip width and the lateral distance from the centerline of the current strip to the centerline of the previous strip.

[0105] Specifically, let the lateral distance from the center line of the current strip to the center line of the previous strip be... Let the effective strip width model be...

[0106] ;

[0107] in, The distance between the walls is [missing information]. Angle of incidence The half-power beamwidth is measured from a planar target at the calibration site.

[0108] The strip overlap is determined according to the following formula. :

[0109] .

[0110] Strip overlap Indicates complete coverage, strip overlap. A value close to 0 indicates almost no overlap.

[0111] 4. Point cloud sparsity estimation: Count the points within a unit surface arc length, normalize them according to the grid area to obtain the statistical point cloud density, and determine the point cloud sparsity based on the ratio of the statistical point cloud density to the reference density.

[0112] Specifically, let the point cloud count within a unit surface arc length be... According to grid area Normalization yields statistical point cloud density Define sparsity .

[0113] in, The reference density obtained for calibration is derived from point count statistics at the reference speed and scan rate. Point cloud sparsity. The larger the value, the sparser the point cloud.

[0114] After calculating the signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity, the signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity are normalized. The normalized signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity are used as components of the imaging quality vector. All components are aligned at the same timestamp within the current period, and the imaging quality vector is output.

[0115] As one possible implementation, the quality assessment module 101 is used to normalize the signal-to-noise ratio, phase noise, stripe overlap, and point cloud sparsity in the following ways:

[0116] According to the family of monotonically differentiable functions ;

[0117] The original mass vector Mapping to the [0,1] interval yields the normalized image quality vector. ;

[0118] in, For signal-to-noise ratio, For phase noise, For strip overlap, Point cloud sparsity; Original mass vector The first in i One portion, Normalized image quality vector The i-th component in Original mass vector The i The normalization function for each component; This is the adjustment coefficient of the signal-to-noise ratio normalization function. This serves as the baseline value for the signal-to-noise ratio. This is the adjustment coefficient for the phase noise normalization function. This is the reference value for phase noise.

[0119] Among them, the signal-to-noise ratio component The larger the better, phase noise component The larger the value, the worse the band overlap component. The larger the better, point cloud sparsity component The bigger, the worse.

[0120] The constraint mapping module 102 is used to generate cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients based on the imaging quality vector through a mapping function.

[0121] In this embodiment of the disclosure, the constraint mapping module 102 transforms the imaging quality vector into an optimized digital scale. The mapping function remains continuous and differentiable to avoid control jitter.

[0122] Image quality vector Mapped to four types of control elements: cost weight Soft constraint penalty coefficient Hard constraint threshold Safety boundary tightening coefficient .set up A family of component-differentiable mappings:

[0123] .

[0124] Typical relationships include: when the sparsity components of a point cloud Increase the speed deviation weight during ascent. And tighten to maximum speed When the phase noise component During ascent, increase the attitude angular velocity weight. With angular acceleration weight And tighten its upper limit; when the strip overlap component Increase lateral deviation penalty during descent. And tighten the maximum step distance When the signal-to-noise ratio component When descending, increase the wall distance deviation weight. And tighten the observation distance range .

[0125] Figure 3 This embodiment of the present disclosure provides a mapping relationship diagram between imaging quality and constraints and cost weights. Taking the imaging quality vector as input, the constraint mapping module 102 generates cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients through a mapping function, which are then input into the rolling optimization module 103 for the quality-aware objective function, geometric consistency, and coverage integrity constraints.

[0126] The generation of cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients will be explained in detail below.

[0127] As one possible implementation, the constraint mapping module 102 is used to generate cost weights in the following manner:

[0128] According to the cost weight mapping formula Determine the cost weights ;

[0129] in, For the imaging quality vector, For benchmark weights, Here is the gain matrix. As the lower bound of the cost weight, The upper limit of the cost weight; cost weight The components include at least one of the following: position tracking weight, velocity weight, acceleration weight, angular velocity weight, angular acceleration weight, lateral deviation weight, wall distance deviation weight, phase noise exceeding limit weight, and point cloud density insufficient weight.

[0130] As one possible implementation, the cost-weight mapping formula satisfies at least one of the following rules:

[0131] Imaging quality vector The sparsity component of point cloud When the cost weight increases, velocity weights in With acceleration weight improve;

[0132] Imaging quality vector Phase noise components When the cost weight increases, Angular velocity weights in With angular acceleration weight improve;

[0133] Imaging quality vector The strip overlap component When reduced, cost weight Lateral deviation weight in improve;

[0134] Imaging quality vector Signal-to-noise ratio components in When reduced, cost weight Wall distance deviation weight improve.

[0135] As one possible implementation, the constraint mapping module 102 is used to generate hard constraint thresholds in the following manner:

[0136] Using a linear tightening formula Determine the hard constraint threshold ;

[0137] in, K is the initial hard constraint threshold, and K is the safety boundary tightening coefficient;

[0138] This is a quality degradation characterization vector. , Image quality vector The signal-to-noise ratio component in Image quality vector Phase noise components in Image quality vector The strip overlap component in the data. Image quality vector The point cloud sparsity component;

[0139] Hard constraint threshold The components include the upper speed limit. Upper limit of angular velocity , acceleration upper limit Minimum wall distance Maximum wall distance Minimum effective point cloud density Maximum step distance At least one of them;

[0140] The linear tightening formula satisfies the rule that when the imaging quality deteriorates, the feasible boundary of the key flight parameters of the UAV shrinks.

[0141] As one possible implementation, the constraint mapping module 102 is used to generate the safety boundary tightening coefficient in the following manner:

[0142] when When the value is less than the first segment threshold, the safety boundary tightening coefficient K adopts the first value;

[0143] when When the threshold of the first segment is exceeded but the threshold of the second segment is not exceeded, the quality degradation mode is triggered, and the safety boundary tightening coefficient K adopts the second value.

[0144] when When the second segment threshold is exceeded, the disturbance over-limit mode is triggered, and the safety boundary tightening coefficient K adopts the third value;

[0145] The second value is greater than the first value, and the third value is greater than the second value.

[0146] Thus, when the infinite norm of the quality degradation characterization vector... Exceeding the first segment threshold With the second segment threshold At that time, the quality degradation mode and the disturbance over-limit mode are triggered respectively, and the hard constraint threshold is adjusted. The tightening is done in segments to prioritize imaging stability and safety.

[0147] As one possible implementation, the constraint mapping module 102 is used to generate soft constraint penalty coefficients in the following manner:

[0148] According to the soft constraint penalty mapping formula Determine the soft constraint penalty coefficient ;

[0149] in, Here, H represents the initial soft constraint penalty coefficient, and H is the adjustment matrix. This is a quality degradation characterization vector;

[0150] Soft constraint penalty coefficient The components include at least one of the following: wall distance deviation from obstacle weight, lateral deviation weight, insufficient point cloud density weight, and phase noise exceeding limit weight;

[0151] The soft constraint penalty mapping formula satisfies the rule that when the imaging quality deteriorates, the constraint stiffness of the soft constraint penalty should be increased.

[0152] In this embodiment, soft constraint penalty coefficients are configured for coverage indicators that are difficult to rigidly meet, serving as barrier terms in the quality-perceived cost function. The soft constraint penalty coefficients are adaptively adjusted according to the degree of image quality degradation, maintaining a relatively smooth soft constraint when image quality is normal and increasing constraint stiffness when image quality deteriorates.

[0153] Initial soft constraint penalty coefficient The adjustment matrix H is a system hyperparameter, which is adjusted according to the actual performance of the application. In practice, it can be preset. =0, H is the identity matrix.

[0154] The rolling optimization module 103 is used to construct a quality-aware cost function based on the aircraft dynamics model, cost weights, soft constraint penalty coefficients, hard constraint thresholds and safety boundary tightening coefficients, and combine geometric consistency and coverage integrity constraints to solve for the optimal control sequence in the finite time domain.

[0155] Among them, the cost weight is used to adjust the importance ratio of the trajectory tracking term and the control smoothing term in the quality perception cost function; the soft constraint penalty coefficient is used to control the penalty intensity of the obstacle term in the quality perception cost function; the hard constraint threshold is used to limit the feasible boundary of the key parameters of UAV flight, which include at least one of velocity, angular velocity, acceleration, wall distance, point cloud density and strip step size; the safety boundary tightening coefficient is used to dynamically narrow the feasible domain of the hard constraint threshold when the imaging quality deteriorates.

[0156] In this embodiment of the disclosure, the rolling optimization module 103 uses a discrete dynamics model, quality-aware cost, and geometric coverage constraints to form a quadratic or nonlinear programming problem, and solves the optimal control of quality-awareness within a finite time domain.

[0157] In each control cycle, the currently estimated imaging quality vector is used. Update cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients. This forms the above quadratic or nonlinear programming problem, and the optimal control sequence is obtained by solving it. Only the first control quantity was issued. The process proceeds to the propulsion and gimbal actuators; new echo and wall distance data are collected, surface geometry and mass vector are updated, and the above process is repeated continuously.

[0158] For solving the optimal control sequence and timing: a prediction domain is used. With step size In each control cycle Construct the problem and call the fast QP (Quadratic Programming) or (SQPSequential Quadratic Programming) solver to obtain the optimal control sequence. Only the first control variable is executed. The remaining control variables are recalculated in the next cycle. The optimization time limit is set as a fixed proportion of the control cycle (commonly set to 30%~60%), and the remaining time is used for data acquisition and preprocessing.

[0159] As one possible implementation, the rolling optimization module 103 is used to construct the quality-aware cost function in the following way:

[0160] Adopting a linearizable aircraft dynamics model Construct a quality-aware cost function

[0161] ;

[0162] The prediction domain is the discrete time domain, and the step size is... The prediction domain length is N ;

[0163] For the first k The state at each time step , For the first k The location of the drone at each time step. For the first k The drone speed at each time step For the first k Attitude parameters or small-angle attitude vectors at each time step;

[0164] For the first k Control quantity per time step , For the first k The expected body acceleration at each time step For the first k The gimbal angular velocity at each time step;

[0165] For the first k The reference trajectory position at each time step For location tracking weights;

[0166] For the first k Reference trajectory velocity at each time step Speed ​​weights;

[0167] For the first k The angular velocity of the drone at each time step As the weight for angular velocity, For acceleration weights;

[0168] For the first k The distance between the drone and the wall of the measured curved surface at each time step. For the desired observation distance, As the weight for wall distance deviation, The smooth hinge penalty function for wall distance deviation;

[0169] For the first k Local lateral step size of adjacent stripes at each time step. For the maximum step size, , The smooth hinge penalty function for lateral deviation;

[0170] For the first k Effective point cloud density at each time step To achieve the minimum effective point cloud density, The point cloud density is insufficient for weighting. For insufficient point cloud density, a smooth hinge penalty function is used.

[0171] For the first k Phase noise proxy at each time step The maximum phase noise threshold. For phase noise exceeding the limit weight, The smooth hinge penalty function is used to compensate for excessive phase noise.

[0172] Position tracking weights Speed ​​weight Angular velocity weight Acceleration weight Cost weight configuration output by constraint mapping module 102;

[0173] Wall distance deviation weight Horizontal deviation weight Insufficient point cloud density weight Phase noise exceeding limits weight The penalty coefficient for soft constraints is determined by the output of constraint mapping module 102;

[0174] Maximum step distance Maximum phase noise threshold Minimum effective point cloud density The hard constraint threshold and safety boundary tightening coefficient are determined by the output of the constraint mapping module 102.

[0175] For example, the aircraft dynamics model is shown below:

[0176] ;

[0177] It is approximately linear over a small angular range, and linearized online at the operating point if necessary.

[0178] The coverage plan provides a reference trajectory location. With the desired visual axis direction .

[0179] The boiler surface is fitted using millimeter-wave point clouds, employing a local spline or tube bank parametric model. Signed distance is defined. ,in For the point set of the surface, Location of the drone. For curved surface collectors The point in the middle. The normal is obtained from the directions of the principal components in the local neighborhood.

[0180] For an approximate model of effective point cloud density, satisfying that the sampling density along the path increases as the speed decreases, we can take... ;in, This is the proportionality coefficient. The scanning frequency of the sensor. Angle of incidence The distance between the walls is [missing information]. For example, point cloud influence functions related to wall distance and incident angle, , For distance attenuation correction, It was obtained by regression analysis of density curves at different distances.

[0181] It is a phase noise surrogate quantity, which is related to angular acceleration and incident geometry.

[0182] in, For a continuously differentiable smooth hinge penalty function, such as γ is the shape parameter, a system hyperparameter used to adjust the smooth transition of constraints from differentiable soft constraints to hard constraints. To balance numerical stability and response sensitivity to defaults, the shape parameter γ of the smooth hinge function is chosen to be a fixed constant of 10.

[0183] formula In It can be , , , .

[0184] As one possible implementation, the rolling optimization module 103 is used for:

[0185] The following constraints are added as geometric consistency and coverage integrity constraints:

[0186] ;

[0187] Wherein, the normal of the measured surface is The sensor's line of sight is , The minimum stripe overlap threshold. For effective strip width, Angle of incidence;

[0188] Minimum wall distance Maximum wall distance Maximum incident angle threshold Minimum strip overlap threshold Maximum step distance Minimum effective point cloud density Speed ​​limit Upper limit of angular velocity , acceleration upper limit The hard constraint threshold and safety boundary tightening coefficient output by the constraint mapping module 102 are updated in real time.

[0189] In this embodiment of the disclosure, the above-mentioned constraints in the rolling optimization module 103 are used to ensure observation consistency and coverage integrity.

[0190] The execution interface module 104 is used to convert the first control quantity in the optimal control sequence into an executable drive command and send it to the UAV propulsion system and gimbal.

[0191] In this embodiment of the present disclosure, the execution interface module 104 solves the optimal control sequence from the rolling optimization module 103. The first control quantity in = It becomes an executable drive command. The propulsion side will... This is translated into desired attitude and motor commands, and then tracked in a closed loop within the flight control circuit. The gimbal side will... The angular velocity commands are converted into gimbal servo commands and aligned with the radar transmit / receive timing. The module records the execution results and errors, and writes them back to the state estimation and the reference trajectory for the next cycle.

[0192] Figure 4 This disclosure provides a flight control and rolling execution flowchart, illustrating the sequential relationship between state estimation, predictive modeling, optimization solution, application of initial control variables, and time-domain rolling.

[0193] like Figure 4As shown, the data transmitted from the millimeter-wave radar and gimbal 201, and the IMU and wall distance sensor 202 are used to obtain an imaging quality vector after quality assessment. The imaging quality vector is then mapped by the constraint mapping module 102 to generate cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients, which are then combined with the UAV status. Reference trajectory position Construct the quality-aware cost function, geometric consistency, and coverage integrity constraints, solve the MPC (e.g., call the MPC), and obtain the optimal control sequence. The execution interface module 104 will execute the optimal control sequence. The first control quantity in The data is converted into executable drive commands to drive the drone's propulsion system and gimbal movements; then, a new round of echo and wall distance data is collected, and the above process is repeated continuously.

[0194] For the timing and edge-end coordination of this system: the system adopts fixed clock alignment. Sensor sampling, mass calculation, mapping update, solution, and command issuance are executed sequentially within one cycle. Millimeter-wave preprocessing and mass calculation are completed at the edge units, and optimization solutions also run at the edge side. The flight control side only receives attitude and velocity expectations and completes high-speed inner-loop control.

[0195] Each module in this system runs in software form on the edge computing unit, or is implemented in a hybrid architecture of programmable logic and general-purpose processors.

[0196] When the program runs on the processor, it executes the above steps to complete quality assessment, constraint mapping, rolling optimization, and instruction issuance. The program is stored in a non-volatile computer-readable medium. Without changing the core idea of ​​this invention, indices such as fringe contrast and registration residuals can be introduced, and normalized mapping can be added. The optimization solver can be implemented differently according to the platform's computing power. Geometric operators can be replaced with explicit pipe array parameter coordinates to ensure consistency with the surface mesh coverage method.

[0197] Feasibility and Adaptability: This system does not depend on any specific machine model or load. Quality normalization and mapping functions... It can be calibrated offline according to different frequency bands, beams, and sampling mechanisms; the point cloud density and strip width models can be fitted by regression from target experiments; the MPC can use a quadratic programming solver or a nonlinear solver, and can be deployed at the edge when real-time requirements are met. The instructions embedded in the storage medium can enable a general-purpose processor or SoC to execute the above steps to complete the invention.

[0198] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes in the drawings are not necessarily essential for implementing this disclosure.

[0199] The sequence numbers of the embodiments disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0200] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A flight control system with imaging quality constraints, characterized in that, The system includes: a quality assessment module, a constraint mapping module, a rolling optimization module, and an execution interface module; The quality assessment module is used to generate an imaging quality vector that includes signal-to-noise ratio, phase noise, strip overlap and point cloud sparsity based on millimeter-wave echo and point cloud data. The constraint mapping module is used to map the cost weights according to the cost weight formula. Determine the cost weights ;in, For the imaging quality vector, For benchmark weights, Here is the gain matrix. As the lower bound of the cost weight, The upper limit of the cost weight; cost weight The components include position tracking weight, velocity weight, acceleration weight, angular velocity weight, angular acceleration weight, lateral deviation weight, wall distance deviation weight, phase noise exceeding limit weight, and point cloud density deficiency weight; the cost weight mapping formula satisfies the following rule: imaging quality vector The sparsity component of point cloud When the cost weight increases, velocity weights in With acceleration weight Improve; Imaging quality vector Phase noise components When the cost weight increases, Angular velocity weights in With angular acceleration weight Improve; Imaging quality vector strip overlap component When reduced, cost weight Lateral deviation weight in Improve; Imaging quality vector Signal-to-noise ratio components in When reduced, cost weight Wall distance deviation weight Improvement; based on the soft constraint penalty mapping formula Determine the soft constraint penalty coefficient ;in, Here, H represents the initial soft constraint penalty coefficient, and H is the adjustment matrix. This is a quality degradation characterization vector. , Image quality vector The signal-to-noise ratio component in Image quality vector Phase noise components in Image quality vector The strip overlap component in the data. Image quality vector The point cloud sparsity components; soft constraint penalty coefficient The components include wall distance deviation from obstacle weight, lateral deviation weight, insufficient point cloud density weight, and phase noise exceeding limit weight; the soft constraint penalty mapping formula satisfies the rule that when the imaging quality deteriorates, the constraint stiffness of the soft constraint penalty is increased; based on the imaging quality vector, a hard constraint threshold and a safety boundary tightening coefficient are generated through a mapping function; the hard constraint threshold is used to limit the feasible boundary of key UAV flight parameters, including velocity, angular velocity, acceleration, wall distance, point cloud density, and strip step size; the safety boundary tightening coefficient is used to dynamically reduce the feasible domain of the hard constraint threshold when the imaging quality deteriorates; The rolling optimization module is used to optimize the aircraft dynamics model. The cost weights, soft constraint penalty coefficients, hard constraint thresholds, and safety boundary tightening coefficients are used to construct the quality-aware cost function. Combining geometric consistency and coverage integrity constraints, the optimal control sequence is obtained in the finite time domain; where the prediction domain is the discrete time domain and the step size is... The prediction domain length is N ; For the first k The state at each time step , For the first k The location of the drone at each time step. For the first k The drone speed at each time step For the first k Attitude parameters or small-angle attitude vectors at each time step; For the first k Control quantity per time step , For the first k The expected body acceleration at each time step For the first k The gimbal angular velocity at each time step; For the first k The reference trajectory position at each time step. For location tracking weights; For the first k Reference trajectory velocity at each time step Speed ​​weights; For the first k The angular velocity of the drone at each time step As the weight for angular velocity, For acceleration weights; For the first k The distance between the drone and the wall of the measured curved surface at each time step. For the desired observation distance, As the weight for wall distance deviation, The smooth hinge penalty function for wall distance deviation; For the first k Local lateral step size of adjacent stripes at each time step. For the maximum step size, , The smooth hinge penalty function for lateral deviation; For the first k Effective point cloud density at each time step To achieve the minimum effective point cloud density, The point cloud density is insufficient for weighting. For insufficient point cloud density, a smooth hinge penalty function is used. For the first k Phase noise proxy at each time step The maximum phase noise threshold. For phase noise exceeding the limit weight, Smooth hinge penalty function for phase noise exceeding limits; position tracking weights Speed ​​weight Angular velocity weight Acceleration weight Cost weight configuration output by the constraint mapping module; wall distance deviation weight Horizontal deviation weight Insufficient point cloud density weight Phase noise exceeding limits weight The maximum step size is determined by the soft constraint penalty coefficient output by the constraint mapping module. Maximum phase noise threshold Minimum effective point cloud density The hard constraint threshold and safety boundary tightening coefficient output by the constraint mapping module are used to determine the safety boundary. The execution interface module is used to convert the first control quantity in the optimal control sequence into an executable drive command and send it to the UAV propulsion system and gimbal.

2. The system according to claim 1, characterized in that, The quality assessment module is used for: The signal-to-noise ratio is calculated by the ratio of the power of the main lobe sample set after the distance-directed FFT to the power of the noise window set. The standard deviation of the demeaned phase sequence of complex echoes of K consecutive pulses within the same line of sight is calculated after unwrapping, and this standard deviation is determined as the phase noise. The effective strip width is determined based on the wall distance, incident angle, and half-power beamwidth; the strip overlap is calculated based on the effective strip width and the lateral distance from the centerline of the current strip to the centerline of the previous strip. The point cloud is counted within a unit surface arc length, and the statistical point cloud density is obtained by normalizing it according to the grid area. The point cloud sparsity is determined based on the ratio of the statistical point cloud density to the reference density. The signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity are normalized. The normalized signal-to-noise ratio, phase noise, strip overlap, and point cloud sparsity are used as components of the imaging quality vector, and the imaging quality vector is output.

3. The system according to claim 2, characterized in that, The quality assessment module is used for: According to the family of monotonically differentiable functions ; The original mass vector Mapping to the [0,1] interval yields the normalized image quality vector. ; in, For signal-to-noise ratio, For phase noise, For strip overlap, Point cloud sparsity; Original mass vector The first in i One portion, Normalized image quality vector The i-th component in Original mass vector The i The normalization function for each component; This is the adjustment coefficient of the signal-to-noise ratio normalization function. This serves as the baseline value for the signal-to-noise ratio. This is the adjustment coefficient for the phase noise normalization function. This is the reference value for phase noise.

4. The system according to claim 1, characterized in that, The constraint mapping module is used for: Using a linear tightening formula Determine the hard constraint threshold ; in, K is the initial hard constraint threshold, and K is the safety boundary tightening coefficient; Hard constraint threshold The components include the upper speed limit. Upper limit of angular velocity , acceleration upper limit Minimum wall distance Maximum wall distance Minimum effective point cloud density Maximum step distance ; The linear tightening formula satisfies the rule that when the imaging quality deteriorates, the feasible boundary of the key flight parameters of the UAV shrinks.

5. The system according to claim 4, characterized in that, The constraint mapping module is used for: when When the value is less than the first segment threshold, the safety boundary tightening coefficient K adopts the first value; when When the threshold of the first segment is exceeded but the threshold of the second segment is not exceeded, the quality degradation mode is triggered, and the safety boundary tightening coefficient K adopts the second value. when When the second segment threshold is exceeded, the disturbance over-limit mode is triggered, and the safety boundary tightening coefficient K adopts the third value; The second value is greater than the first value, and the third value is greater than the second value.

6. The system according to claim 1, characterized in that, The rolling optimization module is used for: The following constraints are added as geometric consistency and coverage integrity constraints: ； Wherein, the normal of the measured surface is The sensor's line of sight is , The minimum stripe overlap threshold. For effective strip width, Angle of incidence; Minimum wall distance Maximum wall distance Maximum incident angle threshold Minimum strip overlap threshold Maximum step distance Minimum effective point cloud density Speed ​​limit Upper limit of angular velocity , acceleration upper limit The hard constraint threshold and safety boundary tightening coefficient output by the constraint mapping module are updated in real time.

Images