Accurate spraying control method and device based on combination of laser scanning and neural network

Abstract

The invention relates to the technical field of spraying control, and discloses a precise spraying control method based on combination of laser scanning and a neural network, and the method comprises the following steps: scanning an environment through a laser radar, dynamically adjusting the scanning harness density based on the real-time speed of equipment, and generating sparse sampling point cloud data; performing compressed sensing reconstruction on the sparse sampling point cloud data, inputting the reconstructed three-dimensional point cloud into a neural network model for target positioning, and outputting a three-dimensional coordinate of a target; calculating a spraying error of the current position of the equipment according to the three-dimensional coordinates; the invention further discloses a precise spraying control device based on combination of laser scanning and the neural network. The precise spraying control device comprises a laser scanning module, a data processing module, a control calculation module, an execution mechanism module and a feedback closed loop module. Through laser radar linkage regulation and control, neural network dynamic allocation and fractional order control, spraying precision improvement, response acceleration and resource optimization are realized.

Description

Technical Field

[0001] The present invention relates to the field of spray control technology, and in particular to a precise spray control method and device based on the combination of laser scanning and neural network. Background Art

[0002] Modern firefighting operations have increasingly stringent requirements for sprinkler accuracy and resource efficiency; traditional sprinkler equipment lacks dynamic environmental perception and adaptive control capabilities, making it difficult to cope with complex terrain distribution.

[0003] Existing technologies use lidar scanning solutions with fixed beam density to stably acquire target coordinates at low speeds, offering simple hardware structures and ease of maintenance. Traditional PID controllers, due to their linear operation characteristics, exhibit good control stability at constant speeds. The independent modular design allows for separate debugging of the positioning, control, and execution systems, reducing initial deployment complexity.

[0004] Existing technologies have significant limitations in dynamic usage scenarios: the fixed scanning mode generates data redundancy or insufficient sampling when the equipment changes speed, resulting in processing delay fluctuations of up to 200%; the traditional PID integer-order model cannot adapt to the nonlinear dynamic characteristics of the spray error, and the control deviation increases by 3 times under rapid speed changes; the timing mismatch between independent modules causes action lag, and the spray landing point deviation exceeds 10cm. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a precise spray control method and device based on the combination of laser scanning and neural network, which solves the problems of inaccurate perception, delayed control response and large spray error of existing spray equipment in dynamic use environment.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a precise spray control method based on the combination of laser scanning and neural network, comprising the following steps: Scan the environment with LiDAR, dynamically adjust the scanning beam density based on the real-time speed of the device, and generate sparse sampling point cloud data; Performing compressed sensing reconstruction on the sparsely sampled point cloud data, inputting the reconstructed three-dimensional point cloud into a neural network model for target positioning, and outputting the three-dimensional coordinates of the target; Calculating the spray error at the current position of the device according to the three-dimensional coordinates, and predicting future acceleration using a random process prediction model based on the time domain rate of change of the spray error, combined with acceleration data collected by an inertial measurement unit and historical motion trajectory; Inputting the predicted future acceleration and spray error into a fractional-order PID controller to generate a sprinkler control instruction; Based on the absolute value of the real-time spray error, a non-smooth correction term is superimposed on the control instruction to output a stability-corrected control signal; The high-frequency piezoelectric nozzle is driven to perform a spraying action according to the corrected control signal, and the spray coverage rate is fed back to the laser radar to dynamically adjust the scanning frequency and sampling rate.

[0007] Preferably, the dynamic adjustment of the scanning beam density: When the device moving speed is greater than a preset first speed threshold, the laser radar scanning beam density is reduced to a first beam density; when the device moving speed is less than the first speed threshold, the scanning beam density is increased to a second beam density, and the beam density adjustment is synchronized with the sampling rate of compressed sensing reconstruction.

[0008] Preferably, the compressed sensing reconstruction includes: An iterative shrinkage threshold algorithm is used to reconstruct sparsely sampled point cloud data, where the measurement matrix dynamically adjusts the sparsity constraint weight according to the device's moving speed; The neural network model is a lightweight PointNet++ architecture, whose input layer receives the reconstructed 3D point cloud coordinates, and the output layer extracts the center of mass position of the target through a maximum pooling operation; The compressed sensing reconstruction and neural network model inference share computing resources. When the point cloud density is lower than a preset threshold, computing power is preferentially allocated to the reconstruction algorithm.

[0009] Preferably, the random process prediction model includes: The Ito process modeling based on equipment acceleration has a drift term that is a linear attenuation function of acceleration and a diffusion term that is a random disturbance term proportional to the square root of acceleration. The spatial sensitivity of the path is quantified by calculating the Malliavin derivative, and the expected acceleration value in the future is predicted by combining the acceleration change trend in the historical motion trajectory. The prediction model is coupled with the input channel of the fractional-order PID controller, and when the predicted acceleration mutation exceeds a preset threshold, an emergency correction mode of the control instruction is automatically triggered.

[0010] Preferably, the fractional-order PID controller includes: The differential operator is defined in first order, and the integral operator is defined in second order; The control parameters are adjusted by frequency distribution algorithm, with phase margin and amplitude margin as optimization targets; The output command of the controller is linked to the kinematic model of the equipment. When the direction of the spray error is opposite to the direction of equipment movement, the weight of the differential term is dynamically increased to suppress overshoot.

[0011] Preferably, the superposition logic of the non-smooth correction term is: When the absolute value of the real-time spray error exceeds a preset first error threshold, a non-smooth compensation term with the same sign as the error is added to the control instruction, and the compensation amplitude is the product of the preset gain coefficient and the absolute value of the error; The activation state of the non-smooth correction term is associated with the device acceleration; The corrected control signal is verified for convergence through the Lyapunov function to ensure that the error decays below the second error threshold within a preset time threshold.

[0012] Preferably, the driving parameters of the high-frequency piezoelectric nozzle include: The pulse width is fixed to a preset time length and is linearly mapped to the driving voltage; The nozzle triggering timing is delayed and compensated according to the predicted acceleration. The compensation time is calculated based on the predicted acceleration and the square of the control period. The array arrangement of the piezoelectric nozzles is orthogonal to the laser scanning direction, and the driving signal of each nozzle is independently adjustable to match the local density distribution of the target.

[0013] Preferably, the spray coverage feedback includes: The environment image after spraying is collected by a multispectral camera, and the sprayed area is segmented using the HSV color space to calculate the coverage percentage. When the coverage rate is lower than the preset coverage rate threshold, the laser radar scanning frequency is increased to the first frequency, and the compressed sensing sampling rate is increased to the first sampling rate; The feedback mechanism is linked to the training data of the neural network model, and the coordinates of the uncovered areas are added to the next round of training data set to optimize positioning accuracy.

[0014] Preferably, the sparsely sampled point cloud data includes: The density range is within the preset density range, the signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, and the point cloud distribution follows the Laplace distribution in the direction of device motion; The sparsity of the point cloud data is dynamically associated with the device speed. When the speed changes, the sparsity is dynamically adjusted according to the rate of change of the device speed. The point cloud attributes are used as prior constraints for compressed sensing reconstruction and for calculating regularization term coefficients in an iterative shrinkage threshold algorithm.

[0015] The present invention also provides a precise spray control device based on the combination of laser scanning and neural network, comprising: Laser scanning module, including a lidar with adjustable beam density and an inertial measurement unit, for real-time collection of 3D point clouds of the environment and device motion data; a data processing module connected to the laser scanning module via a high-speed data interface, comprising an FPGA accelerator and an embedded processor, wherein the FPGA is configured to execute a compressed sensing reconstruction algorithm, and the embedded processor is configured to run a lightweight PointNet++ neural network model; A control calculation module is connected to the data processing module via a PCIe bus and includes a fractional-order PID controller and a non-smooth correction unit. The fractional-order PID controller receives coordinate and device position errors and predicted acceleration to generate initial control instructions. The non-smooth correction unit superimposes a stability compensation term based on an error threshold. The actuator module is connected to the control calculation module through a PWM interface and includes a piezoelectric nozzle array and a high-voltage drive circuit. The trigger delay of the piezoelectric nozzle is less than 15 microseconds. Feedback closed-loop module, the actuator module is configured to transmit the spray coverage data back to the data processing module through the wireless communication module, triggering the dynamic adjustment of the lidar scanning frequency and the compressed sensing sampling rate to form a closed-loop control.

[0016] The present invention provides a precise spray control method and device based on the combination of laser scanning and neural network. Beneficial effects: 1. This invention utilizes a technical solution that combines the coordinated adjustment of LiDAR beam density with device speed, compressed sensing reconstruction, and dynamic allocation of neural network computing power to achieve an optimal balance between point cloud processing efficiency and positioning accuracy. Compared to existing solutions that use fixed scanning parameters, resulting in wasted computing power or insufficient accuracy, this solution solves the problem of balancing efficiency and accuracy.

[0017] 2. By generating acceleration feedforwards through a random process prediction model and integrating the nonlinear dynamic characteristics of a fractional-order PID controller, this technology achieves advanced compensation for spray errors and dynamic delay suppression. Compared to traditional integer-order PID control and static feedforward compensation methods, this overcomes their lag in responding to sudden speed changes.

[0018] 3. This invention employs a technical solution that combines the superposition of non-smooth correction terms triggered by error amplitude with Lyapunov stability verification to ensure control stability under extreme operating conditions. Compared to linear correction or fixed threshold protection strategies, this solution resolves the conflict between slow response and excessive conservatism.

[0019] 4. A technical solution that dynamically adjusts the LiDAR scanning frequency and compressed sensing sampling rate based on spray coverage feedback creates a closed-loop optimization of resource consumption and detection accuracy. Compared to existing open-loop control and fixed-parameter scanning solutions, this overcomes the technical bottleneck of their inability to adapt to operational performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 Schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] Please see the attached Figure 1 The embodiment of the present invention provides a precise spray control method based on the combination of laser scanning and neural network, comprising the following steps: S1. Scan the environment with a lidar, dynamically adjust the scanning beam density based on the real-time speed of the device, and generate sparse sampling point cloud data. The LiDAR scanning module dynamically adjusts the scanning beam density based on the real-time movement speed of the sprinkler system, generating sparsely sampled point cloud data. This process, achieved through a speed-sensing mechanism and coordinated hardware control, resolves the conflict between point cloud data redundancy and processing latency at high speeds while ensuring target positioning accuracy. This dynamic adjustment strategy, combined with subsequent compressed sensing reconstruction and neural network positioning modules, forms a closed data flow loop, providing highly timely input for sprinkler control.

[0023] In some embodiments, the lidar uses an Ouster OS1-64 line scan device with a vertical angular resolution of 0.33° and a horizontal scanning frequency of 10Hz. The inertial measurement unit (IMU) uses an Xsens MTi-670, which collects three-axis acceleration and angular velocity data in real time. The lidar and IMU are fixed via a rigid bracket, ensuring alignment error of less than 2mm.

[0024] Specifically, the LiDAR's scanning beam density is adjusted via the MOS transistor array in the control circuit. The MOS transistor drive signal is dynamically generated by the main control board based on speed data, with a response delay of less than 1ms. The beam shutoff logic follows a "top-down" principle, prioritizing the upper beam emitters while retaining the lower beam for low-profile detection.

[0025] In one possible implementation, the bundle density adjustment rule is defined as: Among them, N lines is the number of activated harnesses, and v is the real-time moving speed of the device, which is measured by the chassis encoder and smoothed by the Kalman filter.

[0026] Specifically, when the speed exceeds a threshold, the lidar scan mode switches to a low-density state, shutting down the top 48 lines of the 48-line beam, leaving only the bottom 16 active. The vertical resolution of the point cloud is reduced to 0.99°, but the horizontal scanning frequency is increased to 20Hz to compensate for the loss of spatial information.

[0027] As an option, the sparsity of the sparsely sampled point cloud data is dynamically related to the device speed. The point cloud density ρ satisfies: Where ρ0=200 points / m3 is the reference density, v is the real-time moving speed of the device, ρ(v) is the dynamically adjusted beam density of the device when the moving speed is v, ΔN=0.15ρ0, v max =5m / s is the maximum design speed of the equipment.

[0028] In some embodiments, the sparsely sampled point cloud data satisfies the following properties: Point cloud distribution: obeys the Laplace distribution in the direction of device movement, and the probability density function is: Among them, p(x, y, z) is the probability density function of the point cloud coordinates in the three-dimensional space, which describes the position distribution characteristics of the device movement direction (x-axis), x-μ x is the absolute deviation between the current x coordinate and the mean, quantifying the position offset, μ x is the mean coordinate value of the equipment moving direction, b = 0.2m is the scale parameter; Signal-to-noise ratio: The original point cloud signal-to-noise ratio (SNR) is greater than 20dB, and the ambient light noise is suppressed by the hardware filtering circuit; Data compression: Run-length encoding (RLE) is used to compress point cloud storage space, with a compression ratio of 6:1.

[0029] In one possible implementation, dynamic bundle density adjustment is linked to the measurement matrix parameters of the compressed sensing reconstruction module. When the bundle density decreases, the sparsity constraint weight λ of the compressed sensing measurement matrix ΦΦ is adjusted synchronously: λ(v) = λ0 + k·v; Among them, λ(v) is the speed-related comprehensive evaluation coefficient, which quantifies the dynamic balance between the equipment movement speed and operation accuracy. λ0 is the static benchmark evaluation coefficient, which characterizes the basic accuracy weight of the equipment when it is stationary or at low speed. k is the speed influence gain coefficient, which describes the linear contribution strength of speed to the comprehensive evaluation value. v is the real-time movement speed of the equipment.

[0030] In some embodiments, the LiDAR scan parameters are optimized through feedback using the following mechanisms: Data quality monitoring: When the reconstruction error exceeds the threshold When forced to increase the bundle density to 64 lines; Energy consumption balancing: Dynamically adjusts the scanning frequency according to the remaining battery power. When the battery power is less than 30%, it switches to energy-saving mode and the beam density is fixed at 32 lines.

[0031] Speed ​​acquisition: The encoder pulse signal is captured by the STM32F407 main control board to calculate the real-time speed v; Threshold judgment: Compare v with the 1.5m / s threshold to generate the MOS tube control signal; Harness switching: The MOS tube array turns off / on the corresponding laser emitter, and the switching time is less than 2ms; Data verification: Verify the integrity of the generated point cloud data. When the packet loss rate exceeds 5%, the retransmission mechanism is triggered.

[0032] S2. Perform compressed sensing reconstruction on the sparsely sampled point cloud data, input the reconstructed three-dimensional point cloud into the neural network model for target positioning, and output the three-dimensional coordinates of the target; The compressed sensing reconstruction module receives sparsely sampled point cloud data generated by dynamically adjusting beam density, recovers the complete 3D point cloud through an iterative optimization algorithm, and inputs the reconstruction results into a lightweight neural network model to output the target's 3D coordinates. This process is linked to the laser scanning module's speed perception mechanism, balancing reconstruction accuracy and real-time performance through a dynamic hardware resource allocation strategy, providing high-precision positioning input for subsequent spray control.

[0033] In some embodiments, compressed sensing reconstruction uses an iterative shrinkage threshold algorithm (ISTA), whose mathematical expression is: Among them, y represents sparse sampling point cloud data, Φ is the measurement matrix, S λ / L is the soft threshold function, λ is the sparsity constraint coefficient, x (k) is the signal estimation value obtained at the kth iteration, representing the target sparse signal in compressed sensing reconstruction. The transposed matrix of the measurement matrix Φ, L is the step size parameter of the gradient descent algorithm, which is associated with the spectral norm of the measurement matrix, y is the sparsely sampled observation data, generated by the lidar beam dynamic adjustment module, and k is the iteration number index.

[0034] In one possible implementation, the construction rules of the measurement matrix Φ include: When the laser radar beam density is 16 lines, Φ uses a partial Fourier matrix whose row vectors are randomly drawn from the standard Fourier basis; When the line density is 64 lines, Φ switches to a Gaussian random matrix.

[0035] Specifically, the sparsity constraint coefficient λ is dynamically associated with the device movement speed v, satisfying: Among them, λ(v) is the speed-related comprehensive evaluation coefficient, which quantifies the dynamic balance between the equipment moving speed and operation accuracy, v is the real-time moving speed of the equipment, and v max Measured by the encoder and processed by sliding average filtering.

[0036] In some embodiments, the neural network adopts an improved PointNet++ architecture, whose structural optimization includes: Layer compression: Delete the third-level set abstraction (SA) layer of the original network, retain the first two SA layers and the feature propagation (FP) layer; Channel number adjustment: The number of output channels of the first-level SA layer is set to 64, and the number of output channels of the second-level SA layer is compressed to 128; Input normalization: The reconstructed point cloud coordinates (x, y, z) are normalized to the range [-1, 1] by the maximum and minimum values.

[0037] In one possible implementation, model training data is generated in the following way: Simulation data: 3D models of corn and wheat were constructed using Blender, with random plant heights (0.5-2.0m) and spacing (0.3-0.8m). Data augmentation: adding Gaussian noise and random occlusion (maximum occlusion rate 30%); Loss function: weighted centroid loss Where c is the center of mass coordinate, and IoU is the intersection over union (IoU) between the predicted and real point clouds.

[0038] As an option, neural network inference is accelerated through the following techniques: Model quantization: convert floating-point weights to INT8 format, and use piecewise linear approximation for activation functions; Memory optimization: Use the TensorRT engine to achieve layer fusion and reduce the number of GPU memory swaps; Output post-processing: coarse-grained centroid coordinates c output by the network coarse Perform Gaussian weighted optimization: Among them, c final is the target coordinate after Gaussian weighted optimization, which is more accurate than the coarse-grained coordinate directly output by the network. coarse is the coarse-grained target coordinate output by the neural network, which serves as the initial center of weighted optimization. i is the point cloud coordinate, and K is the maximum number of cluster points after point cloud clustering.

[0039] S3. Calculate the spray error at the current position of the device based on the three-dimensional coordinates, based on the time domain rate of change of the spray error, combined with the acceleration data collected by the inertial measurement unit and the historical motion trajectory; Based on the target's three-dimensional coordinates output by the neural network, the spray error at the device's current position is calculated in real time. Combined with acceleration data collected by the inertial measurement unit (IMU) and historical motion trajectory, future acceleration is predicted using a stochastic process model. This process addresses control delay issues during high-speed movement by analyzing the error rate of change in the time domain and quantifying path-space sensitivity, providing forward-looking input for the fractional-order PID controller.

[0040] The spray error e(t) is defined as the target centroid coordinate c target With the current nozzle position of the device p current The Euclidean distance of: e(t)=||c target -p current ||2; Among them, e(t) is the spray error, c target is the target center of mass coordinate, p current The nozzle position.

[0041] Device current location p current The positioning accuracy is less than 2cm obtained through the GNSS module.

[0042] In one possible implementation, the time domain error change rate Calculated by first-order differences: Where, e(t) is the spray error, is the instantaneous rate of change of the error signal, which characterizes the rate of change of the error in the control system over time. e(t-Δt) is the historical error of the previous control cycle, which is stored in the ring buffer. Δt is the time step of the control system, that is, the interval between two adjacent error samples.

[0043] Specifically, the random process model uses the Ito diffusion process to describe the dynamic characteristics of equipment acceleration, and its differential equation is expressed as: da(t)=μ(a)dt+σ(a)dW t ; Where da(t) is the instantaneous change in acceleration of the device, which characterizes the random dynamic characteristics of acceleration; a(t) is the acceleration of the device; μ(a) is the drift function, which determines the deterministic evolution trend of acceleration; W t is the Wiener process, σ(a) is the diffusion term coefficient, which quantifies the intensity of random perturbations on acceleration, dW t is the differential of the Wiener process (standard Brownian motion).

[0044] In some embodiments, the path spatial sensitivity index D is calculated based on the Malliavin derivative s a(t), used to correct the prediction results: Among them, D s a(t) is the path spatial sensitivity index, a(t+δt) is the predicted acceleration value at the future time t+δt, a(t-δt) is the measured acceleration value at the historical time t-δt, and δt is the differential time interval, which controls the calculation resolution of the sensitivity index.

[0045] S4, input the predicted future acceleration and spray error into the fractional-order PID controller to generate the nozzle control instruction; the fractional-order PID controller receives the future acceleration from the prediction module The nozzle control instructions are generated through fractional calculus operations. This process combines the device kinematic model with the error dynamics to achieve nonlinear delay compensation and stability control, ensuring that the nozzle movement accurately matches the device motion trajectory.

[0046] In this embodiment, the mathematical expression of the fractional-order PID controller is: in, represents a fractional integral operator of order α, is a fractional differential operator of order β, K p , K i , K d are the proportional, integral, and differential gain coefficients respectively, u(t) is the output command of the controller, and e(t) is the real-time control error, which is defined as the deviation between the set value and the feedback value.

[0047] Specifically, the fractional calculus operator is implemented by a time-domain approximation algorithm: Integral operator: The Riemann-Liouville definition is used, and after discretization, the weighted accumulation of historical errors is calculated using the Grünwald-Letnikov difference formula; Differential operator: Adopt Caputo definition and implement it by convolving the error change rate within a short time window with the exponential decay kernel function.

[0048] Controller parameters (K p , K i , K d , α, β) are adjusted by frequency distribution algorithm. The specific steps include: Frequency domain response test: inject a sweep frequency signal into the device to collect nozzle position response data; Phase margin optimization: Adjust parameters so that the phase margin of the open-loop transfer function at the cutoff frequency is greater than a preset threshold; Robustness verification: Introducing disturbance signals to verify the tolerance of parameters to model uncertainty.

[0049] The controller output instructions are dynamically linked with the equipment kinematic model. The specific implementation method is as follows: Direction correlation compensation: When the spray error direction is opposite to the movement direction of the equipment, the weight of the differential term is increased to suppress overshoot; Acceleration feedforward compensation: The predicted acceleration Inject control instructions as feedforward quantities.

[0050] The controller's output instructions are transmitted to the nozzle drive circuit via the D / A conversion module, while closed-loop stability is ensured through the following mechanisms: Command limit: limit the output voltage within the hardware allowable range to prevent the actuator from saturation; Anti-integral windup: When the error continues to exceed the limit, the integral term accumulation is suspended to avoid control drift; The design of the fractional-order PID controller fully considers the data coupling characteristics with the upstream modules: Predicted acceleration fusion: the output of the prediction module The phase lag caused by the inertia of the equipment is compensated by injection through the feedforward channel; Error dynamic characteristics adaptation: the fractional-order calculus operator is more adaptable to the non-exponential convergence characteristics of the spray error compared to the integer-order PID; Hardware resource collaboration: The control algorithm is implemented in parallel on the FPGA, and the calculation cycle is strictly synchronized with the data acquisition cycle.

[0051] S5. Based on the absolute value of the real-time spray error, a non-smooth correction term is added to the control instruction, and a stability-corrected control signal is output; The initial control command output by the fractional-order PID controller is then superimposed with a nonsmooth correction term based on the absolute value of the real-time spray error to generate a final control signal with optimized stability. This process, through an error amplitude triggering mechanism and a dynamic gain adjustment strategy, addresses the hysteresis of traditional linear corrections to sudden errors. It also achieves closed-loop adaptation with the fractional-order logic of the upstream control module and the physical constraints of the downstream actuator.

[0052] In some embodiments, the activation condition for the non-smooth correction term is strongly correlated with the absolute value of the real-time spray error and the device's motion state. When the absolute value of the spray error exceeds a preset threshold, a compensation value with the same sign as the error is injected into the control command. The magnitude of the compensation value is determined by the product of the error magnitude and the dynamic gain coefficient.

[0053] Specifically, the stacking logic of modifiers includes the following core rules: Threshold trigger mechanism: Set the absolute value threshold of the error to 0.01 meters. When the real-time error exceeds this threshold, the non-smooth correction term is activated; Compensation with the same sign: the direction of the correction term is always consistent with the direction of the error, that is, the compensation amount is positive when the error is positive, and negative when the error is negative; Dynamic gain adjustment: The basic gain coefficient is set to 0.05 V·s / m. When the device acceleration exceeds 3 m / s2, the gain coefficient is increased to 0.08 V·s / m.

[0054] In one possible implementation, dynamic adjustment of the gain coefficient is linked to the device's acceleration and error rate of change. Acceleration data is collected in real time by an inertial measurement unit and filtered using a sliding window to eliminate high-frequency noise. The error rate of change is calculated by taking the weighted average of the error differentials over the last three control cycles, with weights of 0.5, 0.3, and 0.2.

[0055] Preferably, when the device is in a state of rapid acceleration or deceleration, an additional feedforward compensation factor is introduced. This factor is proportional to the square root of the predicted acceleration and is used to compensate for insufficient correction caused by inertia lag.

[0056] In some embodiments, the corrected control signal is verified for convergence using Lyapunov stability theory. An energy function is constructed that includes a combination of an error term and its rate of change. By determining whether the time derivative of the energy function is negative definite, the system is ensured to converge to an equilibrium state within a finite time.

[0057] Specifically, when the energy function derivative does not meet the negative definite condition, the following protection mechanism is triggered: Force the gain coefficient to be reduced to a safe threshold; Temporarily freeze the integral term operation to prevent error accumulation; Send a speed reduction request to the upstream path planning module until stability is restored.

[0058] In this embodiment, the modified control signal works in conjunction with other modules of the system in the following manner: Command limit protection: The final output signal is limited to the voltage tolerance range of the nozzle drive circuit to prevent hardware overload; Timestamp synchronization: The correction term superposition process is strictly synchronized with the controller operation cycle, and the timing deviation is less than 50 microseconds; Feedback closed-loop linkage: When the correction item is triggered continuously for more than the preset number of times, a parameter reset instruction is sent to the laser scanning module to recalibrate the target positioning reference.

[0059] The superposition logic of non-smooth correction terms deeply couples the output characteristics of the upstream control module with the physical limitations of the downstream actuator: Fractional-order PID compatibility: The correction term is added after the controller completes the fractional-order calculus operation to avoid interference with the long memory effect of the fractional-order operator; Nozzle drive adaptation: The voltage amplitude of the correction amount matches the response slope of the nozzle piezoelectric ceramic to prevent mechanical resonance caused by step jump; Real-time performance guarantee: The correction term calculation cycle is compressed to less than 10% of the main control cycle, and parallel processing is achieved through hardware acceleration circuits.

[0060] The non-smooth correction term superposition process sets up multiple anomaly detection mechanisms: Error mutation protection: When the error change rate exceeds the critical value, the correction term superposition is suspended and the control mode is switched to pure proportional control mode; Hardware status monitoring: The temperature and current data of the nozzle drive circuit are read in real time. If an overload risk is detected, the correction amount is attenuated according to the exponential curve; Data integrity check: Perform CRC check on the received control instructions. If the check fails, the data of the previous cycle will be used as the replacement.

[0061] S6. Drive the high-frequency piezoelectric nozzle to perform the spraying action according to the corrected control signal, and feed back the spray coverage rate to the laser radar to dynamically adjust the scanning frequency and sampling rate; The stability-corrected control signal drives the high-frequency piezoelectric nozzle array to execute the spraying action. A multispectral imaging module collects real-time spray coverage data and feeds it back to the LiDAR, dynamically adjusting the scanning frequency and sampling rate. This process forms a closed-loop control architecture that solves the problems of spray blind zone compensation and resource optimization, while also deeply coupling with the command generation logic of the upstream control module and the physical characteristics of the downstream actuator.

[0062] The driving signal generation of the piezoelectric nozzle follows the following rules: Fixed pulse width: The pulse width of the control signal is set to 10 microseconds, and the rise time is less than 0.5 microseconds, matching the mechanical response characteristics of piezoelectric ceramics; Delay compensation: The nozzle triggering timing is feed-forward compensated according to the predicted acceleration. The compensation time is equal to the product of 0.5 times the predicted acceleration and the square of the control period.

[0063] Specifically, when the device is in an accelerating state, the nozzle triggering time is advanced to compensate for the position difference caused by inertia.

[0064] In one possible implementation, the piezoelectric nozzle array is arranged in a direction orthogonal to the laser radar scanning direction, and each nozzle receives a control signal independently. The nozzle drive voltage is dynamically allocated according to the local point cloud density: When the point cloud density is higher than 150 points / m3, the corresponding nozzle drive voltage is increased to 8V; When the point cloud density is lower than 50 points / m3, the driving voltage is reduced to below 3V; The voltage gradient between adjacent nozzles is limited to 5 volts per second.

[0065] The spray coverage is calculated by collecting the post-spray environment image with a multispectral camera: Image preprocessing: extract near-infrared band images and use Gaussian filtering to eliminate light interference; Region segmentation: Set the hue threshold range to 100-140 in the HSV color space, the saturation threshold to be greater than 0.3, and the lightness threshold to be greater than 0.4; Coverage calculation: Counts the percentage of pixels that meet the threshold. When the coverage falls below 85%, parameter adjustment is triggered.

[0066] Preferably, the multispectral camera is rigidly connected to the nozzle array, and the optical field of view covers more than 120% of the spraying area to ensure the integrity of edge area detection.

[0067] The adjustment rules for the LiDAR scanning frequency and compressed sensing sampling rate include: Frequency boost: When coverage falls below a threshold, the lidar scanning frequency is increased from 10 Hz to 20 Hz, and the compressed sensing sampling rate is increased from 50% to 80%; Resource optimization: In low-density areas (coverage greater than 95%), it switches to energy-saving mode, reducing the scanning frequency to 5 Hz and the sampling rate to 30%; Data closure: The coordinate information of uncovered areas is added to the neural network training dataset, and the target positioning accuracy of the next cycle is optimized through incremental learning.

[0068] Specifically, the nozzle drive circuit has multiple built-in protection mechanisms: Temperature monitoring: When the piezoelectric ceramic temperature exceeds 60 degrees Celsius, the frequency reduction operation mode is triggered and the drive frequency is reduced by 50%; Current limiting: Real-time monitoring of the drive current. If it exceeds 2 amperes for three consecutive cycles, the power supply of the corresponding nozzle is cut off and a fault code is reported; Waterproof design: The circuit board is sprayed with three-conformal paint, with a protection level of IP67, suitable for high humidity environments.

[0069] In this embodiment, the actuator and other system modules work together in the following ways: Timing synchronization: The nozzle trigger signal is aligned with the lidar scanning cycle, and the deviation is controlled within 1 millisecond; Data verification: The feedback coverage data is accompanied by timestamps and device location information, and is spatially aligned with the point cloud data; Resource preemption: When the coverage detection module issues an emergency adjustment instruction, the lidar immediately interrupts the current scanning task and prioritizes high-frequency sampling.

[0070] The nozzle drive and feedback adjustment mechanism fully adapts to the overall system architecture characteristics: Control signal compatibility: The corrected control signal voltage range matches the input impedance of the drive circuit to avoid signal reflection loss; Real-time performance guarantee: The nozzle response delay is less than 15 microseconds, ensuring synchronization with the motion trajectory of high-speed mobile devices; Energy efficiency optimization: Dynamically adjust the scanning frequency to reduce lidar power consumption by 40%, extending outdoor operation time.

[0071] The precise spray control device based on the combination of laser scanning and neural network described below and the precise spray control method based on the combination of laser scanning and neural network described above can refer to each other.

[0072] Please see the attached Figure 2 The present invention also provides a precise spray control device based on the combination of laser scanning and neural network, including: Laser scanning module, including a lidar with adjustable beam density and an inertial measurement unit, for real-time collection of 3D point clouds of the environment and device motion data; The data processing module is connected to the laser scanning module through a high-speed data interface and includes an FPGA accelerator and an embedded processor. The FPGA is configured to execute the compressed sensing reconstruction algorithm, and the embedded processor is configured to run the lightweight PointNet++ neural network model. The control calculation module is connected to the data processing module via the PCIe bus and includes a fractional-order PID controller and a non-smooth correction unit. The fractional-order PID controller receives the coordinate and device position errors and the predicted acceleration to generate initial control instructions. The non-smooth correction unit superimposes a stability compensation term based on the error threshold. The actuator module is connected to the control calculation module through a PWM interface and includes a piezoelectric nozzle array and a high-voltage drive circuit. The triggering delay of the piezoelectric nozzle is less than 15 microseconds. The feedback closed-loop module and the actuator module are configured to transmit the spray coverage data back to the data processing module through the wireless communication module, triggering the dynamic adjustment of the lidar scanning frequency and the compressed sensing sampling rate to form a closed-loop control.

[0073] The device of this embodiment can be used to execute the above method embodiment, and its principles and technical effects are similar, so they will not be repeated here.

[0074] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A precise spray control method based on the combination of laser scanning and neural network, characterized in that: The following steps are involved: Scan the environment with LiDAR, dynamically adjust the scanning beam density based on the real-time speed of the device, and generate sparse sampling point cloud data; Performing compressed sensing reconstruction on the sparsely sampled point cloud data, inputting the reconstructed three-dimensional point cloud into a neural network model for target positioning, and outputting the three-dimensional coordinates of the target; Calculating the spray error at the current position of the device according to the three-dimensional coordinates, and predicting future acceleration using a random process prediction model based on the time domain rate of change of the spray error, combined with acceleration data collected by an inertial measurement unit and historical motion trajectory; Inputting the predicted future acceleration and spray error into a fractional-order PID controller to generate a sprinkler control instruction; Based on the absolute value of the real-time spray error, a non-smooth correction term is superimposed on the control instruction to output a stability-corrected control signal; The high-frequency piezoelectric nozzle is driven to perform a spraying action according to the corrected control signal, and the spray coverage rate is fed back to the laser radar to dynamically adjust the scanning frequency and sampling rate.

2. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The dynamic adjustment of the scanning beam density: When the device moving speed is greater than a preset first speed threshold, reducing the laser radar scanning beam density to a first beam density; When the device moving speed is less than the first speed threshold, the scanning beam density is increased to a second beam density, and the beam density adjustment is synchronously linked with the sampling rate of the compressed sensing reconstruction.

3. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The compressed sensing reconstruction includes: An iterative shrinkage threshold algorithm is used to reconstruct sparsely sampled point cloud data, where the measurement matrix dynamically adjusts the sparsity constraint weights based on the device's movement speed. The neural network model uses a lightweight PointNet++ architecture, with its input layer receiving the reconstructed 3D point cloud coordinates and its output layer extracting the target's center of mass through a max-pooling operation. The compressed sensing reconstruction and neural network model inference share computing resources. When the point cloud density is lower than a preset threshold, computing power is preferentially allocated to the reconstruction algorithm.

4. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The random process prediction model includes: The Ito process modeling based on equipment acceleration has a drift term that is a linear attenuation function of acceleration and a diffusion term that is a random disturbance term proportional to the square root of acceleration. The spatial sensitivity of the path is quantified by calculating the Malliavin derivative, and the expected acceleration value in the future is predicted by combining the acceleration change trend in the historical motion trajectory. The prediction model is coupled with the input channel of the fractional-order PID controller, and when the predicted acceleration mutation exceeds a preset threshold, an emergency correction mode of the control instruction is automatically triggered.

5. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The fractional-order PID controller comprises: The differential operator is defined in first order, and the integral operator is defined in second order; The control parameters are adjusted by frequency distribution algorithm, with phase margin and amplitude margin as optimization targets; The output command of the controller is linked to the kinematic model of the equipment. When the direction of the spray error is opposite to the direction of equipment movement, the weight of the differential term is dynamically increased to suppress overshoot.

6. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The superposition logic of the non-smooth correction term is: When the absolute value of the real-time spray error exceeds a preset first error threshold, a non-smooth compensation term with the same sign as the error is added to the control instruction, and the compensation amplitude is the product of the preset gain coefficient and the absolute value of the error; The activation state of the non-smooth correction term is associated with the device acceleration; The corrected control signal is verified for convergence through the Lyapunov function to ensure that the error decays below the second error threshold within a preset time threshold.

7. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The driving parameters of the high-frequency piezoelectric nozzle include: The pulse width is fixed to a preset time length and is linearly mapped to the driving voltage; The nozzle triggering timing is delayed and compensated according to the predicted acceleration. The compensation time is calculated based on the predicted acceleration and the square of the control period. The array arrangement of the piezoelectric nozzles is orthogonal to the laser scanning direction, and the driving signal of each nozzle is independently adjustable to match the local density distribution of the target.

8. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The spray coverage feedback includes: The environment image after spraying is collected by a multispectral camera, and the sprayed area is segmented using the HSV color space to calculate the coverage percentage. When the coverage rate is lower than the preset coverage rate threshold, the laser radar scanning frequency is increased to the first frequency, and the compressed sensing sampling rate is increased to the first sampling rate; The feedback mechanism is linked to the training data of the neural network model, and the coordinates of the uncovered areas are added to the next round of training data set to optimize positioning accuracy.

9. The precise spray control method based on the combination of laser scanning and neural network according to claim 1 is characterized in that: The sparsely sampled point cloud data includes: The density range is within the preset density range, the signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, and the point cloud distribution follows the Laplace distribution in the direction of device motion; The sparsity of the point cloud data is dynamically associated with the device speed. When the speed changes, the sparsity is dynamically adjusted according to the rate of change of the device speed. The point cloud attributes are used as prior constraints for compressed sensing reconstruction and for calculating regularization term coefficients in an iterative shrinkage threshold algorithm.

10. A precise spray control device based on the combination of laser scanning and neural network, according to any one of claims 1 to 9, wherein the precise spray control method based on the combination of laser scanning and neural network is characterized in that: include: Laser scanning module, including a lidar with adjustable beam density and an inertial measurement unit, for real-time collection of 3D point clouds of the environment and device motion data; a data processing module connected to the laser scanning module via a high-speed data interface, comprising an FPGA accelerator and an embedded processor, wherein the FPGA is configured to execute a compressed sensing reconstruction algorithm, and the embedded processor is configured to run a lightweight PointNet++ neural network model; A control calculation module is connected to the data processing module via a PCIe bus and includes a fractional-order PID controller and a non-smooth correction unit. The fractional-order PID controller receives coordinate and device position errors and predicted acceleration to generate initial control instructions. The non-smooth correction unit superimposes a stability compensation term based on an error threshold. The actuator module is connected to the control calculation module through a PWM interface and includes a piezoelectric nozzle array and a high-voltage drive circuit. The trigger delay of the piezoelectric nozzle is less than 15 microseconds. Feedback closed-loop module, the actuator module is configured to transmit the spray coverage data back to the data processing module through the wireless communication module, triggering the dynamic adjustment of the lidar scanning frequency and the compressed sensing sampling rate to form a closed-loop control.

Images