Monomer walking type flattening and tamping device and construction method thereof

Abstract

The application provides a single walking type flattening and tamping device and a construction method thereof, the front end and both sides of a tamping vehicle are provided with a plurality of first cameras and first laser radars; the tamping device is arranged on the tamping vehicle, third and fourth cameras are arranged at the front and rear positions of the bottom of the tamping vehicle, the third and fourth cameras are directed to the tamping position of the end of the tamping device, a second laser radar is arranged on one side of the third camera, and the second laser radar is used for scanning the material distribution condition of the groove bottom; a second camera is arranged at the middle position of the front of the tamping vehicle, a fifth camera is arranged above the tail of the tamping vehicle, and a positioning device is further arranged on the tamping vehicle. Through the synergistic effect of vision and laser radar, the device can accurately extract the center line of the groove, calculate the material distribution and monitor the tamping compactness, effectively solve the technical pain points such as difficult path planning, low flattening precision and unable to realize real-time feedback of compactness in traditional construction, and significantly improve the engineering automation level and construction quality.

Description

Technical Field

[0001] This invention relates to the field of trenching, and in particular to a single-unit mobile leveling and compaction device and its construction method. Background Technology

[0002] In municipal engineering, pipeline laying, and infrastructure construction, the placement, leveling, and compaction of the underlying sand and gravel after trench excavation are crucial foundational processes. Traditional compaction equipment is mostly manually operated or semi-automatically controlled by remote controls. When operating in narrow and complex trenches, this type of equipment lacks the ability to perceive its surroundings in three dimensions and cannot autonomously identify obstacles, trench boundaries, and the geometry of the passageway. This results in the equipment being unable to autonomously plan a safe route, making it prone to deviations or collisions. It not only heavily relies on the operator's experience, posing significant safety hazards, but also struggles to adapt to the long-distance, complex, and ever-changing trench construction environment.

[0003] In the leveling and compaction stages, existing technologies typically employ a blind operation mode with fixed routes and fixed power. On the one hand, because it is impossible to obtain the true distribution and volumetric characteristics of the sand and gravel at the bottom of the trench before operation, the equipment can only perform mechanical, full-coverage movement, unable to intelligently handle material crests or troughs of material accumulation and fill gaps. This directly leads to local depressions or bulges in the leveled base, seriously affecting the quality of subsequent projects. On the other hand, the initial density of the soil and sand and gravel at the bottom of the trench varies greatly in different areas. Traditional equipment uses constant-power vibrators for compaction, lacking real-time monitoring of vibration feedback from the underlying surface and a density assessment mechanism. This easily leads to local under-compaction causing later settlement, or local over-compaction causing sand and gravel breakage and voids, as well as ineffective wear and tear on the mechanical equipment.

[0004] Furthermore, existing compaction equipment still relies on manual spot checks using handheld measuring tools to inspect surface flatness after a single compaction cycle. The equipment itself lacks the capability for visual inspection of the surface flatness after operation. If slope deviations occur or the equipment tilts due to its own weight, the lack of flexible attitude adjustment mechanisms and closed-loop control algorithms prevents the equipment from adaptively adjusting the spatial angle of the leveling and compaction device. Simultaneously, existing single-machine control systems often employ centralized underlying logic with low data processing bandwidth, making it difficult to support high-frequency data fusion from multiple sensors, complex 3D mapping, and edge-cloud collaborative computing. This results in a persistently low level of automation and intelligence for the overall equipment. Summary of the Invention

[0005] The main objective of this invention is to provide a single-unit walking leveling and compaction device and its construction method. The technical problem solved by this application is that existing trench compaction equipment lacks three-dimensional environmental perception and dynamic servo adjustment capabilities, cannot autonomously plan the optimal path according to the terrain and material distribution, and is difficult to perform closed-loop self-correction of excitation power and spatial attitude based on real-time compaction and flatness feedback, resulting in uneven construction quality and low degree of automation.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a single-unit walking leveling and compaction device, wherein multiple first cameras and first lidars are provided at the front end and both sides of the compaction vehicle; The compaction device is installed on the compaction vehicle. The third and fourth cameras are located at the front and rear of the bottom of the compaction vehicle. The third and fourth cameras face the compaction position at the end of the compaction device. A second lidar is installed on one side of the third camera. The second lidar is used to scan the material distribution at the bottom of the trench. The vehicle is equipped with a second camera located in the center front and a fifth camera located above the rear. The vehicle is also equipped with positioning devices.

[0007] In the preferred embodiment, the frame structure includes an upper frame and a lower frame. The upper frame is a parallelepiped frame structure, and the lower frame is a rectangular frame structure. The lower frame and the upper frame are connected and fixed as an integral frame structure by a hinge axis. The bottom of the frame structure is connected to the chassis. The compaction device is installed on the lower frame. One side of the support frame of the compaction device is hinged to one side of the lower frame, and the other side of the support frame of the compaction device is hinged through one end of the second telescopic cylinder. The other end of the lower frame is hinged to one side of the hinge axis of the lower frame and the second upper frame. The upper rear side of the tamping vehicle's upper frame is equipped with a hydraulic station, and the rear side of the lower frame is equipped with a control cabinet.

[0008] In the preferred embodiment, the transverse moving vehicle body and the support frame are slidably connected laterally, and a first motor is provided on both sides of the transverse moving vehicle body. The gear at the output end of the first motor meshes with the rack on the side of the support frame. A guide telescopic column is provided in the middle of the transverse car body. The telescopic rod of the lower part of the guide telescopic column passes through the transverse car body and is connected to the rotary drive device. The drive shaft of the rotary drive device is connected to the connecting gantry. The connecting gantry is connected to the leveling seat through multiple shock-absorbing blocks. A second gear is provided on the rotating end of the rotary drive device, and a second motor is provided on one side of the lower part of the guide telescopic column. The first gear on the second motor meshes with the second gear. The longitudinal moving car body is also equipped with multiple sixth telescopic cylinders. The telescopic end of the sixth telescopic cylinder passes through the transverse moving car body and is connected to the telescopic rod end of the lower end of the guide telescopic column. The guide telescopic column is a multi-stage telescopic device composed of multiple tubes connected together. The leveling seat is equipped with a vibrator; The bottom of the leveling seat is also equipped with a piezoelectric sensor.

[0009] A construction method for a single-unit mobile leveling and compaction device, the method comprising: S1. Collect three-dimensional feature data of the trench environment ahead, and combine it with positioning and map information to generate the automatic walking route of the compaction vehicle. S2. Scan the distribution of materials at the bottom of the trench and plan the leveling and compaction operation route of the compaction device; S3. Monitor vibration feedback parameters during compaction operations in real time and dynamically adjust excitation power based on compaction assessment results; S4. Perform visual inspection of the flatness of the surface after the operation, and adaptively adjust the spatial angle of the compaction device when slope deviation is identified.

[0010] In the preferred embodiment, the steps for generating the automatic walking route in S1 include: The first camera and the first lidar synchronously acquire trench image data and point cloud data; using a multi-sensor extrinsic parameter calibration and point cloud fusion algorithm, the pixel color information in the image data is projected and mapped to the three-dimensional coordinate system of the point cloud data through the camera intrinsic parameter matrix and extrinsic parameter transformation matrix, giving the discrete point cloud texture and color features, and fusing them to form three-dimensional view data. The 3D modeling module deployed in the control cabinet is invoked. The 3D modeling module is a spatial mapping program built based on the octree voxel grid algorithm. The 3D modeling module takes the 3D view data as the input source, discretizes the continuous 3D space into a set of 3D grids according to the set voxel resolution, and uses the ray casting method to mark the grid nodes in the idle, occupied and unknown states. Based on this, a 3D spatial topology map containing the geometry of the entity in front of the trench and the passage boundary is constructed. The random sampling consensus algorithm is used to calculate the normal vectors and fit the plane of the 3D point set in the 3D spatial topology map. The steps are as follows: three non-collinear points are randomly selected from the 3D point set to calculate the parameters of the initial plane model and the corresponding normal vectors; the orthogonal distance from the remaining points in the 3D point set to the initial plane model is calculated, and points with orthogonal distances less than a set distance threshold are included in the inlier set; the above random selection and calculation process is repeated multiple times, and the optimal plane model with the largest number of inlier sets is retained as the fitted trench bottom area, and the normal vector of the optimal plane model is extracted; based on the fitted trench bottom area, the edge contours on both sides are segmented and extracted, and the morphological skeleton extraction algorithm is applied to calculate the geometric midline coordinate sequence along the longitudinal direction as the trench centerline. At the same time, the vertical elevation difference between the surface grid and the bottom area is calculated along the normal direction of the trench centerline to generate depth distribution data. The algorithm planning module generates an automatic walking route for the compaction vehicle to move autonomously based on trench location, depth distribution data, self-positioning, and pre-loaded map data. The specific implementation steps are as follows: First, the algorithm planning module acquires preloaded map data and projects the vehicle's current positioning coordinates, the calculated trench centerline and depth distribution data into the same global cost map after unifying the coordinate system. Then, it updates the danger zone weights in the global cost map based on the depth distribution data. Secondly, a safe offset distance that conforms to the wheel distance of the compacted vehicle is set on both sides of the center line of the trench, and two smooth reference lines parallel to the trench are generated by combining the B-spline curve fitting algorithm. Next, in the global cost map, starting from its own location coordinates, multiple target waypoints are set along the smooth baseline reference line. The global A-star heuristic search algorithm is used to calculate the global barrier-free path node sequence connecting each target waypoint. The steps are as follows: Initialize the open list and the closed list, add the starting point to the open list and calculate the total value of the starting point. The total value is obtained by adding the actual movement cost from the starting point to the current node and the Manhattan distance heuristic cost from the current node to the target waypoint. Select the node with the lowest total value from the list of nodes to be expanded as the current node. If the current node is the target waypoint, backtrack the parent node to generate a global sequence of accessible path nodes. If the current extended node is not the target waypoint, move it to the closed list, and traverse the neighboring nodes around the current extended node that are not blocked by the global cost map danger zone weight. Calculate the total value of each neighboring node and update the parent node pointing to the current extended node. Add the neighboring nodes to the open list and iterate until the target waypoint is found. Finally, a local dynamic window algorithm is used to sample multiple sets of predicted trajectories in the velocity space based on the kinematic constraint model of the compacted vehicle. The steps are as follows: Based on the current linear velocity, angular velocity, and maximum acceleration / deceleration capability of the vehicle, the dynamic speed window range that can be achieved in the next control cycle is calculated. Discretized velocity sampling is performed within a dynamic velocity window, and multiple simulated predicted trajectories within a set time window are generated by combining kinematic model deduction. A comprehensive evaluation function is designed that includes azimuth deviation evaluation, distance from obstacle safety distance evaluation, and current driving speed evaluation. The multiple simulated predicted trajectories are calculated and scored by combining the real-time environmental boundary in the three-dimensional spatial topology map. The optimal predicted trajectory with the highest comprehensive evaluation score is selected and continuously output as motor control commands. An automatic walking route for the compaction vehicle to move autonomously is calculated and generated on both sides of the trench.

[0011] In the preferred scheme, the planning steps for the leveling and compaction operation route in S2 include: The third camera and the second LiDAR are activated to scan the bottom area of ​​the trench simultaneously. A vision and point cloud registration algorithm based on multi-scale feature fusion is used to extract the edge features of the image obtained by the third camera and the geometric features of the depth point cloud obtained by the second LiDAR. The color and texture information of the image is mapped to the depth point cloud to generate three-dimensional terrain point cloud data containing accurate elevation and boundary information, thereby obtaining the distribution of sand and gravel material and the width features of the trench accurately measured. The path planning software deployed inside the control center reads three-dimensional terrain point cloud data and establishes a local spatial coordinate system at the bottom of the trench. The path planning software uses an inverse distance weighted interpolation algorithm to perform spatial difference calculations on the discrete three-dimensional terrain point cloud data and smoothly reconstruct the continuous elevation surface at the bottom of the trench. By combining the preset target leveling elevation of the trench with the continuous elevation surface, a three-dimensional Boolean subtraction operation is performed to integrally solve the accumulation and pitting volume of sand and gravel in different areas, and the local volume distribution of sand and gravel is accurately calculated. The path planning software aims to level the gravel at the bottom of the trench. Based on the local volume distribution of gravel, it generates a coordinated movement coordinate sequence for the first and second motors, and plans a leveling and compaction route to guide the leveling platform. The steps are as follows: Extract the coordinates of peak and trough regions in the volume distribution that exceed the set threshold, mark the peak regions as material collection points and the trough regions as filling points; based on the spatial topological relationship between the material collection points and the filling points, apply the ant colony optimization algorithm to search for the globally optimal material handling and compaction traversal sequence. By combining the linear translation kinematics equations of the lateral moving vehicle body and the rotational kinematics equations of the rotary drive device, the globally optimal sequence is reverse-engineered into a time and position synchronization sequence of the first motor controlling the lateral movement and the second motor controlling the rotation, generating a cooperative movement coordinate sequence, which ultimately constitutes the leveling and compaction operation route.

[0012] In the preferred embodiment, the steps for the path planning software to generate the cooperative movement coordinate sequence include: A data structure is built in memory to construct a two-dimensional discrete grid map of the trench bottom. The grid resolution of the two-dimensional discrete grid map is set to match the physical size of the leveling base bottom surface. The three-dimensional coordinates of the sand and gravel distribution obtained by scanning are projected onto the grid nodes of the two-dimensional discrete grid map. The average elevation of the corresponding point cloud in the grid node is extracted and subtracted from the target leveling elevation. The difference is calculated as the height weight of the corresponding grid. Different grids are assigned corresponding height weights to indicate the surplus and shortage status. A cover-based full-traversal path search algorithm based on ox-plowing cell decomposition is adopted, with the minimization of height-weighted variance as a constraint, to plan continuous planar traversal nodes in a two-dimensional discrete raster map. The steps are as follows: The two-dimensional discrete raster map is divided into multiple simple rectangular cell regions that do not contain internal obstacles; In each rectangular cell region, a reciprocating cleaning path node is generated. A cost function is introduced to evaluate the height weight gradient between adjacent grids. When the reciprocating cleaning reaches the region of abrupt change in height weight, the algorithm forces priority to connect the peak grid with positive height weight to the valley grid with negative height weight. This minimizes the global height weight variance of the leveling seat during the leveling process. In this way, the cleaning path nodes of each rectangular cell region are connected, and continuous planar traversal nodes are planned. Combining the lifting stroke limit of the guide telescopic column, a Z-axis depth coordinate is added to the continuous planar traversal nodes, and finally the three-dimensional spatial trajectory coordinates are synthesized and sent to the servo controllers of each drive motor. The steps are as follows: Read the real-time elevation data of the location of the nodes in the plane traversal, and calculate the target Z-axis depth coordinates based on the compaction and settlement compression amount preset by the vibrator; By comparing the target Z-axis depth coordinate with the maximum extension and minimum retraction of the guide telescopic column, physical boundary truncation is performed on coordinates that exceed the lifting stroke limit. The processed Z-axis depth coordinate is then matrix-stitched with the X-axis and Y-axis coordinates of the planar traversal nodes to synthesize a three-dimensional spatial trajectory coordinate that includes the spatial three-dimensional position and timestamp. The three-dimensional spatial trajectory coordinate is then parsed into pulse control signals through the underlying fieldbus and directly sent to the first motor, the second motor, and the servo controller of the control unit that controls the extension and retraction of the guide telescopic column to execute high-frequency actions.

[0013] In the preferred embodiment, the step of dynamically adjusting the excitation power in S3 includes: When the leveling platform performs compaction work, the piezoelectric sensor collects real-time vibration feedback data of the ground surface at high frequency. The steps are as follows: the piezoelectric ceramic induction array built into the piezoelectric sensor generates micro-charge signals under the alternating load of the vibrator and the ground reaction force; the micro-charge signals are impedance transformed and voltage amplified by the charge amplifier on the hardware acquisition board, and input to the high-frequency analog-to-digital converter for continuous discretization sampling; the high-frequency analog-to-digital converter converts the continuous analog voltage into a digital waveform sequence containing high-precision timestamps and vibration acceleration amplitude according to the set Nyquist sampling frequency, forming real-time vibration feedback data, and transmits it to the signal analysis module in the industrial control computer without delay through the industrial Ethernet bus; The signal analysis module within the industrial control computer filters and performs fast Fourier transform on the real-time vibration feedback data to extract feedback energy characteristics and determine whether the current area's compaction meets the standard. The steps are as follows: First, the signal analysis module uses an adaptive Kalman bandpass filter to denoise the digital waveform sequence, filtering out high-frequency resonance from the mechanical body and low-frequency environmental interference noise. Then, a Hanning window function is added to the denoised time-domain digital waveform sequence to reduce spectral leakage. Subsequently, a fast Fourier transform algorithm is called to convert the time-domain waveform sequence into a frequency-domain energy spectrum. The peak amplitude of the excitation fundamental frequency and the amplitude components of the second and third harmonics are extracted from the frequency-domain energy spectrum. A continuous compaction evaluation algorithm is applied to calculate the ratio of the sum of multiple harmonic amplitudes to the excitation fundamental frequency amplitude as the compaction monitoring value, i.e., the extracted feedback energy characteristics. Finally, the calculated compaction monitoring value is compared differentially with the preset target soil compaction characteristic threshold in the system database to determine whether the current area's compaction meets the standard. When the compaction level is found to be below standard, the power control program calculates a dynamic compensation coefficient and increases the output power of the vibrator in real time until the compaction monitoring result meets the standard requirements. The steps are as follows: If the current compaction monitoring value is lower than the target soil compaction characteristic threshold, the power control program extracts the numerical difference between the two as the control deviation; a fuzzy proportional-integral-derivative control algorithm is introduced, and the control deviation and the rate of change of the deviation are used as input variables for fuzzification processing. Fuzzy inference is performed in combination with a preset expert fuzzy control rule table, and the dynamic compensation coefficient is output after defuzzification; the power control program converts the dynamic compensation coefficient into a pulse width modulation duty cycle command or a hydraulic flow servo valve opening control command, and increases the driving frequency or eccentric torque amplitude of the vibrator in real time, thereby improving the overall output power; during the power increase and compaction operation extension process, the piezoelectric sensor continuously performs high-frequency acquisition actions in a loop, and the signal analysis module and the power control program perform high-speed iterative calculations synchronously to form a closed-loop feedback adjustment mechanism until the compaction monitoring result reaches and stabilizes within the standard requirements range.

[0014] In the preferred embodiment, the step of adaptively adjusting the spatial angle in S4 includes: The fourth camera continuously captures images of the compaction position at the end of the compaction device. The steps are as follows: The fourth camera is equipped with a global shutter and automatic exposure control logic to continuously capture high-resolution images of the ground area behind the leveling seat that has just completed the compaction operation at a set frame rate; during the image capture process, the current timestamp of the system and the three-dimensional spatial coordinates of the compaction device are recorded simultaneously to form a continuous image sequence with pose and time label; then, the continuous image sequence with pose and time label is transmitted losslessly through a high-speed industrial communication interface and input into the data buffer of the machine vision inspection module deployed in the industrial control computer. The machine vision detection module uses an image edge feature extraction algorithm to visually detect whether there is unevenness on the ground surface. The steps are as follows: The machine vision detection module extracts a continuous image sequence from the data buffer. First, it uses a Gaussian blur filtering algorithm to smooth the image matrix to remove high-frequency noise interference. Subsequently, the Canny edge detection algorithm was used to calculate the gradient magnitude and gradient direction of the pixel gray values ​​in the image matrix, and non-maximum suppression and double threshold edge connection were performed to extract the set of pixels at the edge of the land surface contour. By combining the Hough linear transform algorithm, the extracted set of pixels on the surface contour edge is mapped to the parameter space, and the physical contour of the actual surface is detected and fitted. Extract the tilt angle and flatness features of the physical contour of the straight line, compare the tilt angle of the physical contour of the straight line with the preset horizontal reference threshold, and if the tilt angle difference exceeds the tolerance range of the horizontal reference threshold, visual detection will detect the unevenness of the ground surface due to slope deviation. If a slope deviation is detected visually, the attitude control module calculates the angle compensation variable and outputs an analog signal to adjust the extension or retraction of the second telescopic cylinder, so that the compaction device produces a corresponding angle change to conform to the horizontal reference plane. The steps are as follows: the attitude control module extracts the spatial angle between the actual physical contour of the straight line on the ground and the horizontal reference plane as the initial error value. The PID closed-loop control algorithm is invoked, and the initial error value is input into the proportional-integral-derivative controller. Combined with historical attitude feedback data, the angle compensation variable used to eliminate slope deviation is calculated. Based on the inverse kinematics model of the overall suspension structure of the compaction device, the angle compensation variable is nonlinearly transformed into the target linear displacement parameter of the second telescopic cylinder. The digital-to-analog conversion unit in the attitude control module converts the target linear displacement parameter into a continuous voltage analog signal or current analog signal, which is then used to drive the hydraulic proportional valve through a servo amplifier. The hydraulic proportional valve adjusts the flow rate and direction of the hydraulic oil injected into the second telescopic cylinder in real time according to the analog signal, and accurately executes the extension or retraction of the second telescopic cylinder. This drives the support frame of the compaction device to rotate around the hinge axis of the lower frame, so that the compaction device produces a precise corresponding angle change to dynamically fit the horizontal reference surface, thereby completing the closed-loop automatic correction of the leveling surface.

[0015] In the preferred embodiment, the various data processing and control steps in the method are deployed and executed through an intelligent control system software mounted on an onboard industrial control computer. The detailed deployment and execution steps include: The intelligent control system software adopts a modular microservice architecture based on the ROS robot operating system, including hardware driver nodes, environment perception and mapping nodes, path planning nodes, density analysis nodes, and attitude servo control nodes. The steps are as follows: compile and deploy the core communication master node of the ROS robot operating system at the Linux operating system level of the vehicle-mounted industrial control computer; construct independent microservice processes by writing code, configuring the hardware driver node as a topic publisher responsible for parsing sensor-level messages and encapsulating actuator instructions; configuring the environment perception and mapping node as a topic subscriber and publisher that fuses multi-source data and generates a 3D spatial topology map; configuring the path planning node, density analysis node, and attitude servo control node as independent service processing units that execute corresponding core algorithms; and achieving decoupled asynchronous data interaction and distributed parallel computing between the microservice processes through the ROS topic publish-subscribe mechanism, service call mechanism, and action communication mechanism. The vehicle-mounted industrial control computer establishes low-level communication with the first camera, first lidar, third camera, second lidar, piezoelectric sensor, and fourth camera via a field Ethernet bus. The steps are as follows: configure the TCP / IP network protocol stack and UDP multicast protocol in the vehicle-mounted industrial control computer; assign fixed static IP addresses to the first camera, first lidar, third camera, second lidar, piezoelectric sensor, and fourth camera through a gigabit industrial switch in the field Ethernet bus, and establish a Socket long connection socket at the application layer; the hardware driver node reads the raw data stream from each sensor port by calling the corresponding driver API interface and using a multi-threaded polling mechanism, serializes and packages the continuous raw data stream according to the ROS standard sensor message format and synchronizes the global timestamp, thus completing high-bandwidth, low-latency low-level data acquisition and communication interaction. A closed-loop control flow is formed, from data acquisition and edge computing to hardware action execution. The steps are as follows: the bottom-level sensors continuously acquire data on the trench environment characteristics and the machine's operating status. The hardware driver node publishes the serialized and packaged standard messages to the ROS core communication bus network at high frequency. The environmental perception and mapping node, path planning node, density analysis node, and attitude servo control node subscribe to the required feature messages in real time. Utilizing the edge computing power of the on-board industrial control computer, combined with the wireless communication network, the cloud computing cluster is called to perform large-scale point cloud registration and model rendering calculations. After the calculation is completed, the optimal planned trajectory, excitation power dynamic compensation coefficient, and attitude angle compensation variables are generated and published to the bottom-level action execution node. The bottom-level action execution node reverse-engineers the received control variables into bottom-level pulse signals or analog drive commands to control the first motor, second motor, exciter, and second telescopic cylinder, and sends them to the servo drivers of each hardware mechanism to execute specific physical actions. The environmental and state changes caused by the actions of the hardware mechanism are captured again by the sensor array, seamlessly entering the next control cycle, thereby constructing a closed-loop control flow for the whole machine with high dynamic response and real-time self-correction.

[0016] This invention provides a single-unit, mobile leveling and compaction device and its construction method. This application utilizes multi-sensor front fusion and 3D spatial topology mapping technology to endow the compaction device with full-scene perception capabilities in complex trench environments. Combining global A-satellite heuristic search and local dynamic window algorithm, the device can autonomously avoid dangerous areas and generate smooth automatic walking routes, completely eliminating the limitations of manual driving or remote control, and greatly improving the safety and autonomous robustness of trench operations.

[0017] This application utilizes visual and lidar registration to scan the material distribution at the bottom and accurately reconstructs the elevation surface through spatial difference calculation, transforming the leveling operation into an intelligent path optimization process based on volume surplus and deficit. Combined with ox-plowing-style cellular decomposition and ant colony optimization algorithms, the equipment can prioritize and precisely push sand and gravel from higher areas to lower areas, achieving intelligent dynamic material allocation and fundamentally eliminating the localized unevenness caused by traditional blind leveling.

[0018] This application innovatively introduces a high-frequency piezoelectric sensor and a fast Fourier transform algorithm to achieve real-time analysis of the energy characteristics of surface vibration feedback. By establishing a closed-loop mechanism of continuous compaction evaluation and fuzzy proportional-integral-derivative control, the equipment can dynamically increase or decrease the output power of the vibrator in milliseconds according to the actual compaction state of the underlying medium. This variable-power adaptive compaction mode ensures that the compaction degree across the entire area strictly meets the standards while effectively avoiding energy waste and mechanical fatigue caused by over-compaction.

[0019] This application integrates machine vision edge feature extraction and posture inverse calculation technologies, enabling the equipment to have self-inspection and self-healing capabilities for work quality. By continuously capturing images and fitting the straight physical contours of the actual ground surface, once a slope deviation exceeding the tolerance is detected, the system can automatically adjust the telescopic cylinder through a hydraulic proportional valve to flexibly adjust the ground contact angle of the compaction device, achieving high-precision horizontal correction of the working surface.

[0020] This application constructs a modular microservice architecture based on a robot operating system, establishing a complete data link from multi-threaded polling data acquisition by underlying sensors and edge computing power cloud collaboration to hardware actuators. This high-bandwidth, low-latency distributed computing system ensures efficient fusion of multi-source heterogeneous data and real-time solving of complex control algorithms, giving the entire single-unit walking leveling and compaction device extremely high dynamic response capability and system stability. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the construction of the compaction vehicle of this invention; Figure 2 This is a front view structural diagram of the compaction device of the present invention; Figure 3 This is a front view of the compaction device of the present invention; Figure 4 This is a rear view structural diagram of the compaction device of the present invention.

[0022] In the diagram: 1. Compactor; 101. Hydraulic station; 102. Control cabinet; Upper frame 2; Lower frame 3; Track 4; Compactor 5; Guide telescopic column 501; First telescopic cylinder 502; Lateral moving body 503; First motor 504; Support frame 505; Second motor 506; First gear 507; Second gear 508; Connecting gantry 509; Shock absorber 510; Leveling seat 511; Vibrator 512; Rack 513; First camera 6; First lidar 7; Second camera 8; Third camera 9; Fourth camera 10; Fifth camera 11; Second lidar 12; Second telescopic cylinder 13. Detailed Implementation

[0023] Example 1 like Figure 1-4 As shown, a single-unit walking leveling and compaction device is provided with multiple first cameras 6 and first lidar 7 at the front and sides of the compaction vehicle 1. The compaction device 5 is installed on the compaction vehicle 3. The compaction vehicle 3 has a third camera 9 and a fourth camera 10 at the front and rear positions of the bottom. The third camera 9 and the fourth camera 10 face the compaction position at the end of the compaction device 5. A second lidar 12 is provided on one side of the third camera 9. The second lidar 12 is used to scan the material distribution at the bottom of the trench. The compaction vehicle 3 is equipped with a second camera 8 at the front center and a fifth camera 11 at the top rear. The compaction vehicle 3 is also equipped with positioning equipment.

[0024] In the preferred embodiment, the frame structure includes an upper frame 2 and a lower frame 3. The upper frame 2 is a parallelepiped frame structure, and the lower frame 3 is a rectangular frame structure. The lower frame 3 and the upper frame 2 are connected and fixed as an integral frame structure by a hinge axis. The bottom of the frame structure is connected to the chassis 9. The compaction device 5 is installed on the lower frame 3. One side of the support frame 505 of the compaction device 5 is hinged to one side of the lower frame 3, and the other side of the support frame 505 of the compaction device 5 is hinged to one end of the second telescopic cylinder 13. The other end of the lower frame 3 is hinged to the position of the hinge axis of the lower frame 3 and the second upper frame 10. The upper rear side of the upper frame 2 of the compaction vehicle 1 is equipped with a hydraulic station 101, and the rear side of the lower frame 3 is equipped with a control cabinet 102.

[0025] In the preferred embodiment, the transverse vehicle body 503 is slidably connected to the support frame 505 in the transverse direction. The transverse vehicle body 503 is provided with a first motor 504 on both sides. The gear at the output end of the first motor 504 meshes with the rack 513 on the side of the support frame 505. The transverse car body 503 is provided with a guide telescopic column 501 in the middle. The telescopic rod of the lower part of the guide telescopic column 501 passes through the transverse car body 503 and is connected to the rotary drive device. The drive shaft of the rotary drive device is connected to the connecting gantry 509. The connecting gantry 509 is connected to the leveling seat 511 through multiple shock absorbers 510. A second gear 508 is provided on the rotating end of the rotary drive device, and a second motor 506 is provided on one side of the lower part of the guide telescopic column 501. The first gear 507 on the second motor 506 meshes with the second gear 508. The longitudinal moving car body 501 is also equipped with multiple sixth telescopic cylinders 502. The telescopic end of the sixth telescopic cylinder 502 passes through the transverse moving car body 503 and is connected to the telescopic rod end of the lower end of the guide telescopic column 501. The guide telescopic column 501 is a multi-stage telescopic device composed of multi-stage tubes. The leveling seat 511 is equipped with a vibrator 512; The bottom of the leveling seat 511 is also equipped with a piezoelectric sensor 10.

[0026] Example 2 Further explanation in conjunction with Example 1, such as Figure 1-4The structure shown illustrates a construction method for a single-unit, mobile leveling and compaction device, the method comprising: S1. Collect three-dimensional feature data of the trench environment ahead, and combine it with positioning and map information to generate the automatic walking route of the compaction vehicle 1. S2. Scan the distribution of materials at the bottom of the trench and plan the leveling and compaction route of the compaction device 5. S3. Monitor vibration feedback parameters during compaction operations in real time and dynamically adjust excitation power based on compaction assessment results; S4. Perform visual inspection of the flatness of the surface after the operation, and adaptively adjust the spatial angle of the compaction device 5 when slope deviation is identified.

[0027] Example 3 Further explanation in conjunction with Example 1, such as Figure 1-4 The structure shown includes the following steps in generating the automatic walking route in S1: The first camera 6 and the first lidar 7 simultaneously acquire trench image data and point cloud data; using a multi-sensor extrinsic parameter calibration and point cloud fusion algorithm, the pixel color information in the image data is projected and mapped to the three-dimensional coordinate system of the point cloud data through the camera intrinsic parameter matrix and extrinsic parameter transformation matrix, giving the discrete point cloud texture and color features, and fusing them to form three-dimensional view data. The three-dimensional modeling module deployed in the control cabinet 102 is invoked. The three-dimensional modeling module is a spatial mapping program built based on the octree voxel grid algorithm. The three-dimensional modeling module takes the three-dimensional view data as the input source, discretizes the continuous three-dimensional space into a set of three-dimensional grids according to the set voxel resolution, and uses the ray casting method to mark the grid nodes in the idle, occupied and unknown states. Based on this, a three-dimensional spatial topology map containing the geometric structure of the entity in front of the trench and the passage boundary is constructed. The random sampling consensus algorithm is used to calculate the normal vectors and fit the plane of the 3D point set in the 3D spatial topology map. The steps are as follows: three non-collinear points are randomly selected from the 3D point set to calculate the parameters of the initial plane model and the corresponding normal vectors; the orthogonal distance from the remaining points in the 3D point set to the initial plane model is calculated, and points with orthogonal distances less than a set distance threshold are included in the inlier set; the above random selection and calculation process is repeated multiple times, and the optimal plane model with the largest number of inlier sets is retained as the fitted trench bottom area, and the normal vector of the optimal plane model is extracted; based on the fitted trench bottom area, the edge contours on both sides are segmented and extracted, and the morphological skeleton extraction algorithm is applied to calculate the geometric midline coordinate sequence along the longitudinal direction as the trench centerline. At the same time, the vertical elevation difference between the surface grid and the bottom area is calculated along the normal direction of the trench centerline to generate depth distribution data. The algorithm planning module generates an automatic walking route for the compaction vehicle 1 to move autonomously based on the trench location, depth distribution data, self-positioning, and pre-loaded map data. The specific implementation steps are as follows: First, the algorithm planning module acquires the preloaded map data, and projects the current positioning coordinates of the solidified vehicle 1, the calculated center line of the trench, and the depth distribution data into the same global cost map after unifying the coordinate system. The weight of the dangerous area in the global cost map is updated according to the depth distribution data. Secondly, a safe offset distance that conforms to the wheel track of the compaction vehicle is set on both sides of the center line of the trench, and two smooth reference lines parallel to the trench are generated by combining the B-spline curve fitting algorithm. Next, in the global cost map, starting from its own location coordinates, multiple target waypoints are set along the smooth baseline reference line. The global A-star heuristic search algorithm is used to calculate the global barrier-free path node sequence connecting each target waypoint. The steps are as follows: Initialize the open list and the closed list, add the starting point to the open list and calculate the total value of the starting point. The total value is obtained by adding the actual movement cost from the starting point to the current node and the Manhattan distance heuristic cost from the current node to the target waypoint. Select the node with the lowest total value from the list of nodes to be expanded as the current node. If the current node is the target waypoint, backtrack the parent node to generate a global sequence of accessible path nodes. If the current extended node is not the target waypoint, move it to the closed list, and traverse the neighboring nodes around the current extended node that are not blocked by the global cost map danger zone weight. Calculate the total value of each neighboring node and update the parent node pointing to the current extended node. Add the neighboring nodes to the open list and iterate until the target waypoint is found. Finally, the local dynamic window algorithm is used to sample multiple sets of predicted trajectories in the velocity space based on the kinematic constraint model of the solidified vehicle 1. The steps are as follows: Based on the current linear velocity, angular velocity, and maximum acceleration / deceleration capability of vehicle 1, calculate the dynamic speed window range that can be achieved in the next control cycle; Discretized velocity sampling is performed within a dynamic velocity window, and multiple simulated predicted trajectories within a set time window are generated by combining kinematic model deduction. A comprehensive evaluation function is designed that includes azimuth deviation evaluation, distance from obstacle safety distance evaluation, and current driving speed evaluation. The multiple simulated predicted trajectories are calculated and scored by combining the real-time environmental boundary in the three-dimensional spatial topology map. The optimal predicted trajectory with the highest comprehensive evaluation score is selected and continuously output as motor control commands. An automatic walking route for the compaction vehicle 1 to move autonomously is calculated and generated on both sides of the trench.

[0028] In the automatic walking route generation step, trench image data and point cloud data are simultaneously acquired using a first camera and a first LiDAR. To address the limitation of single-sensor perception dimension, the system employs a multi-sensor extrinsic parameter calibration and point cloud fusion algorithm. Specifically, the pixel color information in the image data is projected onto the three-dimensional coordinate system of the point cloud data using a camera intrinsic parameter matrix and an extrinsic parameter transformation matrix. The mathematical expression for this projection mapping relationship is: ; In the formula, Represents the depth value in the camera coordinate system. and Represents the two-dimensional pixel coordinates of the image. The intrinsic parameter matrix representing the camera. This represents the extrinsic parameter transformation matrix between the LiDAR and the camera. , , This represents the three-dimensional coordinates of the original point cloud in the world coordinate system. Through the aforementioned matrix operations, the system can accurately assign realistic texture and color features to the discrete point cloud, fusing them to form high-fidelity three-dimensional view data. The beneficial effect of this processing method is that it greatly enriches the feature dimensions of environmental data, enabling the system to combine geometric depth and visual color for comprehensive judgment. This effectively avoids boundary misidentification caused by the lack of texture features at the edges of grooves or uneven lighting distribution, significantly improving the accuracy and robustness of all-weather environmental perception.

[0029] The system then invokes the 3D modeling module deployed in the control cabinet. This module is a spatial mapping program built based on an octree voxel grid algorithm. It takes 3D view data as input and discretizes the continuous 3D space into a set of 3D grids according to a set voxel resolution. In the spatial state definition stage, the system uses ray casting to simulate the sensor's line of sight, traversing along the line connecting the sensor origin to the 3D point cloud coordinates. The system marks the grid where the ray ends as occupied, the intermediate grids traversed by the ray as idle, and the occluded areas not touched by the ray as unknown. Based on these state markings, the system constructs a 3D spatial topology map containing the geometry of the entity in front of the trench and the passage boundary. The beneficial effect of this step is that the octree data structure can perform a very high proportion of data compression on a large-scale scattered 3D space, automatically filtering redundant dense point clouds, greatly reducing the memory usage and computing power consumption of the on-board industrial control computer, and meeting the needs of real-time mapping in industrial sites. At the same time, it clearly distinguishes between the three physical states of occupied, vacant, and unknown, providing an absolutely safe spatial constraint physical boundary for subsequent barrier-free path planning.

[0030] After acquiring the 3D spatial topology map, the system uses a random sampling consensus algorithm to calculate the normal vectors and fit the plane to the 3D point set in the map. The system first randomly selects three non-collinear points from the 3D point set and calculates an initial plane model parameter and its corresponding normal vector. Let the equation of this initial plane be: ; In the formula, , , This represents the directional components of a plane normal vector in the three dimensions of space. This represents the constant distance from the plane to the origin of the coordinate system. , , This represents the coordinates of any point in space. Next, the system calculates the orthogonal distances from all remaining points in the 3D point set to this initial plane. The mathematical formula for the orthogonal distance is: ; In the formula, The calculated orthogonal distance, , , The system assigns the 3D coordinates of the remaining test points. Points with an orthogonal distance less than a set threshold are grouped into an interior point set. After iterating through the random selection and distance calculation process multiple times, the system selects and retains the optimal plane model with the most points in the interior point set as the fitted trench bottom surface region, and extracts the normal vector of this optimal plane. Based on the fitted trench bottom surface region, the system performs region growth and segmentation on the point cloud, extracting the contour lines of the two side edges. Subsequently, the system applies a morphological skeleton extraction algorithm to calculate a geometric central axis coordinate sequence along the longitudinal direction of the trench as the trench centerline, and simultaneously calculates the vertical elevation difference between the surrounding surface grid and the bottom surface region along the normal direction of the trench centerline, thereby generating trench depth distribution data. The beneficial effect of this step is that the random sampling consensus algorithm has a strong immunity to noise points such as splashed mud and scattered gravel commonly found at construction sites, and can robustly fit the real trench bottom working surface. Combined with the skeleton extraction algorithm, even under extremely irregular and harsh working conditions during trench excavation, the true geometric center line can be accurately located, providing a high-precision absolute reference for the symmetrical compaction operation of the entire compaction equipment.

[0031] The algorithm planning module generates an autonomous driving route for the compactor vehicle based on the acquired trench location, depth distribution data, the compactor vehicle's own positioning, and pre-loaded map data. First, the module acquires the pre-loaded map data and performs a unified coordinate system transformation on the compactor vehicle's current positioning coordinates, the real-time calculated trench centerline, and the depth distribution data, then projects them onto the same global cost map. The system dynamically updates the danger zone weights in the global cost map using the depth distribution data; edge areas with larger depth differences are assigned higher danger weights. Second, the system sets safe offset distances on both sides of the trench centerline, conforming to the physical wheelbase characteristics of the compactor vehicle, and uses a B-spline curve fitting algorithm to generate two smooth reference lines parallel to the trench. The mathematical expression of the B-spline curve is: ; In the formula, This represents the coordinates of the point on the generated smooth curve corresponding to the parameter position. The set of coordinates of the control points representing the curve. The order is represented as B-spline basis functions, This represents the node parameter vector. The beneficial effect of this processing step is that it comprehensively integrates the fall hazard factors caused by static scenes and dynamic elevation differences using a global cost map, establishing a robust safety buffer zone. Simultaneously, the baseline reference line generated by the B-spline curve algorithm possesses excellent second-derivative continuity, completely eliminating sharp inflection points at the junction of straight lines and curves. This ensures a smooth chassis transition for the tracked compactor vehicle when switching between straight and curved travel, significantly reducing mechanical wear and the risk of vehicle rollover and instability caused by track wear.

[0032] After constructing the global cost map, the system uses the vehicle's current self-positioning coordinates as the starting point for pathfinding. Multiple target waypoints are set ahead along the tangent direction of the smooth baseline reference line. The global A* heuristic search algorithm is used to calculate the global barrier-free path node sequence connecting each target waypoint. Specifically, the system initializes an open list and a closed list in memory, adds the starting point to the open list, and calculates the total cost value of that starting point. The formula for evaluating the total cost value is: ; In the formula, Represents the current node in the space Total assessed value This represents moving from the pathfinding starting point to the current node. The accumulated actual movement cost, Represents the current node The Manhattan distance to the target waypoint is used to predict the cost. In each iteration, the system selects the node with the lowest total value from the open list as the current expansion node. If the current expansion node is determined to be the target waypoint, it backtracks through all its parent node relationships, sequentially connecting them to generate a globally unobstructed path node sequence. If the current expansion node is not the target waypoint, it is moved to the closed list to prevent duplicate searches, and all neighboring nodes around the current expansion node that are not blocked by the high-pressure weight of dangerous areas in the global cost map are traversed outwards. The system recalculates the total value of each neighboring node and updates its parent node pointer to point to the current expansion node. Then, these valid neighboring nodes are added to the open list, and the process is repeated iteratively until the target waypoint is finally found. The beneficial effect of the A* algorithm is that it cleverly combines the accurate convergence characteristics of Dijkstra's algorithm with the efficiency of the best-first search algorithm. Relying on the precise guidance of heuristic costs, the system can solve a globally shortest safe path from the starting point to the target point at an extremely fast computing speed, while strictly avoiding high-weight danger areas such as landslides. This eliminates engineering accidents caused by navigation getting stuck in local dead ends and resulting in equipment shutdown from the underlying algorithm logic.

[0033] Ultimately, the system employs a local dynamic window algorithm, sampling multiple predicted trajectories within the velocity closure space based on the kinematic constraint model of the compactor chassis. First, the system rigorously calculates the dynamic velocity window limit that the equipment can reach within the next micro-control cycle, based on the real-time linear velocity, real-time angular velocity read from the compactor at the current moment, and the maximum acceleration / deceleration capability of the motor. Then, within this safe and controlled dynamic velocity window, the system performs equidistant, dispersed velocity sampling, and combines this with the time-step extrapolation using the differential slip kinematic model unique to tracked vehicles, generating multiple simulated predicted trajectories distributed within the set time window. To select the optimal solution from the massive number of trajectories, the system designs a comprehensive evaluation function to score them; the formula for this comprehensive evaluation function is as follows: ; In the formula, Represents sampling linear velocity and sampling angular velocity The final comprehensive evaluation score of the generated single predicted trajectory, An evaluation index representing the azimuth deviation between the orientation of the predicted trajectory's end and the global target's path heading. The evaluation index represents the shortest safe distance between the entire predicted trajectory and the physical boundary of the real-time environment within the three-dimensional spatial topology map. This represents the efficiency evaluation index of driving speed under the current predicted sampling. , , These represent the corresponding weight adjustment coefficients. The system combines the real-time environmental physical boundaries in the 3D spatial topology map to perform parallel calculations and scoring on all generated simulated predicted trajectories. The optimal predicted trajectory with the highest absolute value of the comprehensive evaluation score is selected, and its corresponding linear velocity and angular velocity are converted into drive control commands for the chassis motor and continuously output. This accurately calculates and generates an automatic walking route for the compaction vehicle to move autonomously on both sides of the trench. The significant benefit of the local dynamic window algorithm is that it overcomes the fatal flaw of pure global planning algorithms in dealing with dynamic obstacles such as sudden personnel movement or temporary material spillage at the construction site. By hard-constraining the vehicle's physical limit kinematic parameters into the algorithm's underlying layer, it ensures that every driving command generated by the system is a physical action that the chassis drive mechanism is actually capable of responding to and will never lose control. This achieves a perfect coupling of high-precision macro-global tracking and millisecond-level micro-local dynamic obstacle avoidance when the compaction vehicle is traveling at high speed on the edge of complex and steep trenches.

[0034] Example 4 Further explanation in conjunction with Example 1, such as Figure 1-4 The planning steps for the leveling and compaction route in S2, as shown in the diagram, include: The third camera 9 and the second lidar 12 are activated to simultaneously scan the bottom area of ​​the trench. A vision and point cloud registration algorithm based on multi-scale feature fusion is used to extract the edge features of the image obtained by the third camera 9 and the geometric features of the depth point cloud obtained by the second lidar 12. The color and texture information of the image is mapped to the depth point cloud to generate three-dimensional terrain point cloud data containing accurate elevation and boundary information, thereby obtaining the distribution of sand and gravel material and the width features of the trench that are accurately measured. The path planning software deployed inside the control center reads three-dimensional terrain point cloud data and establishes a local spatial coordinate system at the bottom of the trench. The path planning software uses an inverse distance weighted interpolation algorithm to perform spatial difference calculations on the discrete three-dimensional terrain point cloud data and smoothly reconstruct the continuous elevation surface at the bottom of the trench. By combining the preset target leveling elevation of the trench with the continuous elevation surface, a three-dimensional Boolean subtraction operation is performed to integrally solve the accumulation and pitting volume of sand and gravel in different areas, and the local volume distribution of sand and gravel is accurately calculated. The path planning software aims to level the gravel at the bottom of the trench. Based on the local volume distribution of the gravel, it generates a coordinated movement coordinate sequence for the first motor 504 and the second motor 506, and plans a leveling and compaction route to guide the leveling seat 511. The steps are as follows: Extract the coordinates of peak and trough regions in the volume distribution that exceed the set threshold, mark the peak regions as material collection points and the trough regions as filling points; based on the spatial topological relationship between the material collection points and the filling points, apply the ant colony optimization algorithm to search for the globally optimal material handling and compaction traversal sequence. Combining the linear translation kinematics equations of the lateral moving vehicle body 503 with the rotational kinematics equations of the rotary drive device, the globally optimal sequence is reverse-engineered into a time and position synchronization sequence for the first motor 504 to control lateral movement and the second motor 506 to control rotation, generating a cooperative movement coordinate sequence, which ultimately constitutes the leveling and compaction operation route.

[0035] In the planning steps of leveling and compacting the operation route, the system first activates the third camera and the second LiDAR to simultaneously scan the bottom area of ​​the trench. To achieve high-precision unification of heterogeneous data acquired by different sensors, the system employs a vision-point cloud registration algorithm based on multi-scale feature fusion. This algorithm extracts the edge features of the image acquired by the third camera and the geometric features of the depth point cloud acquired by the second LiDAR, and accurately maps the color and texture information of the image to the depth point cloud through a rigid body transformation matrix. The mathematical expression for point cloud spatial mapping is: ; In the formula, This represents the set of coordinates of the target's 3D point cloud after fusion, which includes color information. This represents the set of original geometric depth point cloud coordinates obtained from the second lidar scan. This represents the multidimensional rotation matrix obtained through calibration. This represents the translation vector matrix between sensors. Through the above matrix operations, the system generates 3D terrain point cloud data containing precise elevation and boundary information, thereby obtaining the distribution of sand and gravel materials and the accurately calculated width characteristics of trenches. The beneficial effect of this step is that it fully leverages the complementary advantages of lidar's strong resistance to ambient light interference and high ranging accuracy, and visual cameras' rich edge details and keen color features. The 3D terrain point cloud data generated by multi-sensor fusion completely eliminates the perception blind spots of a single sensor, providing the system with an initial operating environment map with extremely high resolution and true physical scale, greatly improving the accuracy and reliability of subsequent spatial calculations.

[0036] Subsequently, the path planning software deployed within the control center reads the aforementioned 3D terrain point cloud data and establishes a local spatial coordinate system for the bottom of the trench, using the current center point of the compaction vehicle as the origin. Since the spatial distribution of the scanned point cloud data is often non-uniform and discrete, the path planning software employs an inverse distance weighted interpolation algorithm to perform spatial difference calculations on the discrete 3D terrain point cloud data, smoothly reconstructing the continuous elevation surface of the trench bottom. The calculation formula for the inverse distance weighted interpolation algorithm is: ; In the formula, This represents the estimated elevation of any coordinate point on the continuous elevation surface to be determined. This represents the elevation value of a known sampling point in a 3D terrain point cloud data. This represents the straight-line distance in Euclidean space between the coordinate point to be estimated and the known sampling points. The parameter representing the power-law parameter of the rate at which the control weights decay with distance. This represents the total number of known sampling points around the point that participated in estimating its elevation. The beneficial effect of this processing step is that the inverse distance weighted interpolation algorithm can dynamically allocate weights based on distance, effectively repairing missing holes in the point cloud dataset caused by local occlusion, filtering out individual abrupt noise points, and transforming the disordered discrete data point matrix into a mathematically continuous and smooth three-dimensional surface model. This prevents high-frequency vibrations in mechanical devices when following a rough point cloud trajectory, improving the smoothness of the leveling operation.

[0037] After reconstructing the continuous elevation surface, the system performs a three-dimensional Boolean subtraction operation on the continuous elevation surface, combining the preset target leveling elevation of the trench with the surface elevation, and then solves for the accumulation and pitting volume of sand and gravel in different areas through surface integration. The mathematical expression for solving the local sand and gravel volume distribution by integration is: ; In the formula, This represents the precisely calculated volume difference of local sand and gravel. Positive values ​​represent the peak accumulation area where sand and gravel are abundant, while negative values ​​represent the trough depression area where sand and gravel are scarce. It represents the integral closed region on the two-dimensional plane, that is, the cross-sectional projection range of the trench; This represents a continuous elevation surface function reconstructed using the inverse distance weighted interpolation algorithm; This represents the preset target leveling elevation constant required by the construction process. The beneficial effect of this step is that, through a rigorous calculus mathematical model, the material surplus or shortage situation, which was originally estimated roughly by human observation, is transformed into a high-precision numerical quantitative indicator. This allows the system to accurately grasp the volume of material that needs to be scraped or filled in each local area, providing absolutely accurate data support for subsequent intelligent material allocation and completely eliminating the need for rework caused by blind leveling.

[0038] The path planning software focuses on leveling the sand and gravel at the bottom of the trench as its core optimization objective. Based on precisely calculated local sand and gravel volume distribution, it generates a coordinated movement coordinate sequence for the first and second motors. Specifically, the system extracts the coordinates of peaks and troughs in the volume distribution that exceed a set tolerance threshold. Peaks above the elevation are marked as material collection points, and troughs below the elevation are marked as filling points. Based on the physical topology of the material collection and filling points in the local spatial coordinate system, the system applies an ant colony optimization algorithm to search for the globally optimal material handling and compaction traversal sequence. The formula for calculating the ant state transition probability in the ant colony optimization algorithm is: ; In the formula, Representing the Virtual ants only move from the current material collection point or filling point Move to the next target point The probability magnitude; Representative path arrive The total pheromone concentration trajectory on the track is dynamically updated with the number of iterations; The heuristic information representing path transitions is set as a point. Time The reciprocal of the spatial distance between them; A factor representing the relative importance of pheromone concentration; A factor representing the relative importance of heuristic information; Representing the This algorithm uses only the set of available target points that virtual ants have not yet visited. The beneficial effect of this algorithm is that, for complex, nondeterministic polynomial problems such as multi-point material handling and filling, the ant colony optimization algorithm, thanks to its positive feedback mechanism and distributed computing characteristics, can quickly escape local optima and search for the globally optimal material handling traversal sequence with the shortest total travel path and lowest energy consumption within the vast solution space, greatly improving the construction and operation efficiency of the compaction equipment.

[0039] Finally, the system combines the linear translation kinematics equations of the lateral moving vehicle with the rotational kinematics equations of the rotary drive device to inversely solve the globally optimal sequence found in the search, resulting in a time and position synchronization sequence for the first motor controlling the lateral movement and the second motor controlling the rotation. The core expression for the inverse kinematics solution is: ; ; In the formula, and This represents the set of target space coordinates in the globally optimal traversal order output by the ant colony optimization algorithm. Represented by the kinematic equations of linear translation The first motor obtained by reverse engineering is at time [time]. The target lateral linear displacement; Represented by rotational kinematics equations The second motor obtained by reverse engineering is at time [time]. The target rotational angular displacement is determined. The system generates a coordinated movement coordinate sequence, which ultimately forms the leveling and compaction running path. The beneficial effect of this step is that it directly maps the abstract algorithm path data into a temporal physical quantity that can be recognized by the underlying hardware actuators, ensuring that the first motor controlling the lateral movement and the second motor controlling the rotation achieve sub-millisecond rigid synchronization on the time axis. This effectively avoids mechanical interference or tearing and pulling that may occur when multiple actuators move in complex spatial trajectories, ensuring high-precision tracking of the leveling seat's movement trajectory in three-dimensional space, and ultimately achieving perfect intelligent automatic leveling of the sand and gravel at the bottom of the trench.

[0040] In the preferred embodiment, the steps for the path planning software to generate the cooperative movement coordinate sequence include: A data structure is built in memory to construct a two-dimensional discrete grid map of the trench bottom. The grid resolution of the two-dimensional discrete grid map is set to match the physical size of the bottom surface of the leveling seat 511. The three-dimensional coordinates of the sand and gravel distribution obtained by scanning are projected onto the grid nodes of the two-dimensional discrete grid map. The average elevation of the corresponding point cloud in the grid node is extracted and subtracted from the target leveling elevation. The difference is calculated as the height weight of the corresponding grid. Different grids are assigned corresponding height weights to indicate the surplus and shortage status. A cover-based full-traversal path search algorithm based on ox-plowing cell decomposition is adopted, with the minimization of height-weighted variance as a constraint, to plan continuous planar traversal nodes in a two-dimensional discrete raster map. The steps are as follows: The two-dimensional discrete raster map is divided into multiple simple rectangular cell regions that do not contain internal obstacles; In each rectangular cell region, a reciprocating cleaning path node is generated; a cost function is introduced to evaluate the height weight gradient between adjacent grids. When the reciprocating cleaning reaches the region of abrupt change in height weight, the algorithm forces priority to connect the peak grid with positive height weight to the valley grid with negative height weight, so that the global height weight variance of the leveling seat 511 is dynamically minimized during the leveling process. In this way, the cleaning path nodes of each rectangular cell region are connected, and continuous planar traversal nodes are planned. Based on the lifting stroke limitation of the guide telescopic column 501, a Z-axis depth coordinate is added to the continuous planar traversal nodes, and finally the three-dimensional spatial trajectory coordinates are synthesized and sent to the servo controllers of each drive motor. The steps are as follows: Read the real-time elevation data of the location of the nodes in the plane traversal, and calculate the target Z-axis depth coordinate based on the compaction and settlement compression amount preset by the vibrator 512; The target Z-axis depth coordinate is compared with the maximum extension and minimum retraction of the guide telescopic column 501. The coordinates that exceed the lifting stroke limit are physically truncated. The processed Z-axis depth coordinate is then matrix-stitched with the X-axis and Y-axis coordinates of the planar traversal nodes to synthesize a three-dimensional spatial trajectory coordinate that includes the three-dimensional spatial position and timestamp. The three-dimensional spatial trajectory coordinate is then parsed into pulse control signals through the underlying fieldbus and directly sent to the first motor 504, the second motor 506, and the servo controller of the control unit that controls the extension and retraction of the guide telescopic column 501 to execute high-frequency actions.

[0041] In the step of generating the cooperative movement coordinate sequence in the path planning software, the system first establishes a specific data structure in memory to construct a two-dimensional discrete grid map of the trench bottom. To ensure that the data processing granularity matches the physical actuator, the system sets the grid resolution of this two-dimensional discrete grid map to be completely consistent with the physical dimensions of the leveling base. Subsequently, the system projects the three-dimensional coordinates of the scanned sand and gravel distribution onto the grid nodes of this two-dimensional discrete grid map. The system extracts the average elevation of the corresponding point cloud within the grid node and subtracts the target leveling elevation, calculating the difference as the height weight of the corresponding grid. Different grids are assigned corresponding height weights to indicate the material surplus or shortage status. The mathematical expression for calculating the height weight is: ; In the formula, The row coordinates in a two-dimensional discrete raster map are Column coordinates are The height weight of a specific grid. This represents the total number of 3D point cloud coordinates that fall within the specific grid area. Represents the number of cells that fall within this grid. The actual elevation value of a 3D point cloud. This represents the final target leveling elevation constant required by the construction process. The beneficial effect of this step is that by forcibly binding the high-resolution grid size to the physical dimensions of the leveling base, changes in the data of each grid directly correspond to a real mechanical leveling action, avoiding ineffective waste of computing power. Simultaneously, the height weight formula transforms the massive and chaotic 3D point cloud into a two-dimensional weight matrix that intuitively identifies volume surplus or deficit, providing a small but highly significant input for subsequent algorithms, fully ensuring the real-time performance and reliability of spatial difference calculations on the onboard industrial control computer.

[0042] After acquiring a two-dimensional discrete grid map with marked surplus and deficit states, the system employs a cover-based full-traversal path search algorithm based on ox-plowing cell decomposition, with the minimization of height-weighted variance as the core constraint, to plan continuous planar traversal nodes within the two-dimensional discrete grid map. The system first divides the irregularly shaped two-dimensional discrete grid map into multiple simple rectangular cell regions without internal obstacles, and generates reciprocating cleaning path nodes within each rectangular cell region. To achieve intelligent material allocation, the system introduces a cost function to evaluate the height-weighted gradient between adjacent grids. The mathematical expression of the cost function is: ; In the formula, Represents the leveling seat [cite: 30, 43] from the current grid node. Move to the adjacent target grid node The transfer cost score, Represents the geometric distance of spatial transfer between two nodes. Represents the starting grid node High weight, Represents the target mesh node High weight, and This represents the dimensionless weighting control coefficient. When repeatedly sweeping to regions with abrupt changes in height weight, due to the existence of difference terms with extremely large differences, the aforementioned cost function will drive the algorithm to prioritize connecting peak grids with positive height weights to trough grids with negative height weights. Simultaneously, the system defines the global height weight variance formula as follows: In the formula [cite_start], The global height weight variance representing the current working face. This represents the total number of valid grids in a two-dimensional discrete raster map. Representing the The current height weight of each grid. This forced cross-transfer mechanism dynamically minimizes the global height weight variance during the leveling process. The system uses this logic to connect the cleaning path nodes of each rectangular cell region, ultimately planning a continuous planar traversal node. The beneficial effect of this step is that the ox-plowing cell decomposition method ensures 100% coverage of the work area without dead zones from a geometric topology perspective. Combining the cost function and the constraint of minimizing the global height weight variance, the equipment no longer blindly moves randomly across the entire map, but intelligently pushes excess sand and gravel from the accumulation points directly to the nearest pit, achieving rapid equalization of material volume in physical space with the shortest mechanical movement distance, greatly improving the efficiency of the leveling operation and the final surface flatness.

[0043] After planning the driving trajectory in a two-dimensional plane, the system, considering the physical limitations of the lifting stroke of the guide telescopic column, adds Z-axis depth coordinates to the continuous planar traversal nodes, ultimately synthesizing the three-dimensional spatial trajectory coordinates and sending them to the servo controllers of each drive motor for execution. The system first reads the real-time elevation data of the location of the planar traversal node and calculates the target Z-axis depth coordinates based on the pre-set compaction and settlement compression amount of the vibrator. The formula for calculating the target Z-axis depth coordinates is: ; In the formula, This represents the theoretically required Z-axis depth coordinate of the target. This represents the real-time elevation data of the uncompacted surface at the current location of the traversed node in the plane. This represents the preset compaction, settlement, and compression physical quantity produced by the vibrator acting on a specific soil medium at standard power. After calculating the theoretical coordinates, the system compares the target Z-axis depth coordinates with the maximum elongation and minimum retraction of the guide telescopic column, and performs strict physical boundary truncation on coordinates exceeding the lifting stroke limits. The mathematical piecewise function logical expression for the truncation is as follows: ; In the formula, Represents the final safe Z-axis depth coordinate after boundary safety verification. This represents the maximum absolute extension limit position allowed by the mechanical structure of the guide telescopic column. This represents the minimum permissible absolute retraction limit position. The system performs matrix concatenation of the processed Z-axis depth coordinates with the X-axis and Y-axis coordinates of the planar traversal nodes to synthesize a three-dimensional spatial trajectory coordinate system containing the three-dimensional spatial position and precise timestamp. Finally, the system resolves the three-dimensional spatial trajectory coordinates into pulse control signals recognizable by the servo drive via the underlying fieldbus, and directly sends them to the servo controllers of the first motor, the second motor, and the control unit that controls the extension and retraction of the guide telescopic column to execute high-frequency actions. The beneficial effects of this step are significant. By introducing the compaction settlement compression amount for feedforward compensation calculation, it ensures that the final settlement surface after a single compaction can perfectly fit the engineering target elevation, eliminating the hysteresis elevation error caused by material deformation under stress in conventional equipment. At the same time, the forced physical boundary truncation process constitutes a hard safety defense at the software level, completely eliminating the possibility of the controller issuing dangerous commands exceeding the mechanical physical limits from the algorithmic root, effectively preventing hydraulic cylinder explosion or hard jamming damage to the mechanical structure. The final multi-axis matrix splicing technology ensures absolute time synchronization of the three degrees of freedom of lateral movement, rotation and lifting in high-speed spatial motion, giving large engineering equipment extremely high trajectory tracking accuracy and smoothness of coordinated action.

[0044] Example 5 Further explanation in conjunction with Example 1, such as Figure 1-4 The structure shown, the steps for dynamically adjusting the excitation power in S3 include: When the leveling seat 511 performs compaction work, the piezoelectric sensor 10 collects real-time vibration feedback data of the ground surface at high frequency. The steps are as follows: the piezoelectric ceramic induction array built into the piezoelectric sensor 10 generates micro-charge signals under the alternating load of the vibrator 512 and the ground reaction force; the micro-charge signals are impedance transformed and voltage amplified by the charge amplifier on the hardware acquisition board, and input to the high-frequency analog-to-digital converter for continuous discretization sampling; the high-frequency analog-to-digital converter converts the continuous analog voltage into a digital waveform sequence containing high-precision timestamps and vibration acceleration amplitude according to the set Nyquist sampling frequency, forming real-time vibration feedback data, and transmits it to the signal analysis module in the industrial control computer without delay through the industrial Ethernet bus; The signal analysis module within the industrial control computer filters and performs fast Fourier transform on the real-time vibration feedback data to extract feedback energy characteristics and determine whether the current area's compaction meets the standard. The steps are as follows: First, the signal analysis module uses an adaptive Kalman bandpass filter to denoise the digital waveform sequence, filtering out high-frequency resonance from the mechanical body and low-frequency environmental interference noise. Then, a Hanning window function is added to the denoised time-domain digital waveform sequence to reduce spectral leakage. Subsequently, a fast Fourier transform algorithm is called to convert the time-domain waveform sequence into a frequency-domain energy spectrum. The peak amplitude of the excitation fundamental frequency and the amplitude components of the second and third harmonics are extracted from the frequency-domain energy spectrum. A continuous compaction evaluation algorithm is applied to calculate the ratio of the sum of multiple harmonic amplitudes to the excitation fundamental frequency amplitude as the compaction monitoring value, i.e., the extracted feedback energy characteristics. Finally, the calculated compaction monitoring value is compared differentially with the preset target soil compaction characteristic threshold in the system database to determine whether the current area's compaction meets the standard. When the assessment indicates that the compaction degree does not meet the standard, the power control program calculates the dynamic compensation coefficient and increases the output power of the vibrator 512 in real time until the compaction degree monitoring result meets the standard requirements. The steps are as follows: If the current compaction degree monitoring value is lower than the target soil compaction degree characteristic threshold, the power control program extracts the numerical difference between the two as the control deviation; a fuzzy proportional-integral-derivative control algorithm is introduced, and the control deviation and the rate of change of the deviation are used as input variables for fuzzification processing. Fuzzy inference is performed in combination with the preset expert fuzzy control rule table, and the dynamic compensation coefficient is output after defuzzification; the power control program converts the dynamic compensation coefficient into a pulse width modulation duty cycle command or a hydraulic flow servo valve opening control command, and increases the driving frequency or eccentric torque amplitude of the vibrator 512 in real time, thereby improving the overall output power; during the power increase and compaction operation extension process, the piezoelectric sensor 10 continuously performs high-frequency acquisition actions in a loop, and the signal analysis module and the power control program perform high-speed iterative calculations synchronously to form a closed-loop feedback adjustment mechanism until the compaction degree monitoring result reaches and stabilizes within the standard requirements range.

[0045] In the step of dynamically adjusting the excitation power, when the leveling seat performs compaction work, the piezoelectric sensor collects real-time vibration feedback data of the ground surface at high frequency. Specifically, the piezoelectric ceramic induction array built into the piezoelectric sensor generates a micro-charge signal under the alternating load of the exciter and the ground reaction force. This micro-charge signal is extremely weak and exhibits high impedance characteristics. It undergoes impedance transformation and voltage amplification via a charge amplifier on the hardware acquisition board. The mathematical expression for charge amplification is: ; In the formula, This represents the continuous analog voltage signal output by the charge amplifier. This represents the amount of micro-charge generated by the piezoelectric ceramic induction array when it is subjected to deformation. This represents the feedback capacitor parameter within the charge amplifier's internal circuitry. The amplified analog voltage is then input to a high-frequency analog-to-digital converter (ADC) for continuous discretization sampling. To ensure the digital signal remains undistorted, the ADC operates strictly according to a set Nyquist sampling frequency, calculated mathematically as follows: ; In the formula, This represents the actual sampling frequency of the high-frequency analog-to-digital converter. This represents the highest effective frequency component in the real-time vibration feedback signal. Through the aforementioned high-frequency discretization sampling, continuous analog voltages are converted into digital waveform sequences containing high-precision timestamps and vibration acceleration amplitudes, forming real-time vibration feedback data. This data is then transmitted without delay to the signal analysis module within the industrial control computer via an industrial Ethernet bus. The beneficial effect of this step is that, utilizing the high-frequency response characteristics of piezoelectric ceramics and the high-fidelity conversion of the charge amplification circuit, the system can accurately capture the extremely short and minute changes in surface reaction force during compaction. Simultaneously, adhering to the Nyquist sampling theorem completely eliminates spectral aliasing during digital signal conversion, while the high-bandwidth transmission of the industrial Ethernet bus ensures zero loss and zero delay in the massive amounts of high-frequency sampled data, laying a solid underlying hardware and data foundation for subsequent accurate signal analysis.

[0046] After receiving real-time vibration feedback data, the signal analysis module within the industrial control computer filters and performs a Fast Fourier Transform (FFT) to extract feedback energy characteristics, thereby determining whether the current area's compaction meets the standard. The signal analysis module first uses an adaptive Kalman bandpass filter to denoise the digital waveform sequence, accurately filtering out high-frequency resonance from the mechanical body and low-frequency environmental interference noise. To reduce the spectral leakage effect of discrete signals during time-frequency conversion, the system applies a Hanning window function to the denoised time-domain digital waveform sequence. Subsequently, the system calls the FFT algorithm to convert the time-domain waveform sequence into a frequency-domain energy spectrum. In the frequency-domain energy spectrum, the signal analysis module extracts the peak amplitude of the fundamental frequency of the excitation and the amplitude components of the second and third harmonics. The system applies a continuous compaction evaluation algorithm to calculate the ratio of the sum of multiple harmonic amplitudes to the amplitude of the fundamental frequency of the excitation as the compaction monitoring value. The mathematical formula for this compaction monitoring value is: ; In the formula, CMV represents the extracted feedback energy characteristic, i.e., the density monitoring value. This represents the dimensionless constant coefficients calibrated according to different soil types and mechanical operating conditions. This represents the fundamental frequency amplitude extracted from the frequency domain energy spectrum. The amplitude component representing the second harmonic. This represents the amplitude component of the third harmonic. The system performs a differential comparison between the calculated compaction monitoring value and the preset target soil compaction characteristic threshold in the system database to determine whether the current area's compaction meets the standard. The beneficial effect of this processing step is that, because the compacted material exhibits increasingly stronger nonlinear stiffness characteristics as compaction increases, this surface nonlinear stiffness directly leads to a surge in high-order harmonic energy in the vibration feedback signal. Through Fast Fourier Transform and a rigorous harmonic ratio formula, the system cleverly transforms the difficult-to-measure physical compaction of soil into an energy characteristic index that can be accurately quantified and calculated in the frequency domain. Combined with the Hanning window function and adaptive Kalman filtering, interference from mechanical operating noise is completely eliminated, giving the compaction evaluation results extremely high anti-interference capability and industrial-grade reliability, overcoming the technical challenge of blind spots in traditional compaction operations where the compaction of the underlying soil layer cannot be visually identified.

[0047] When the assessment indicates that the current compaction level is below the standard, the power control program begins calculating the dynamic compensation coefficient, increasing the exciter's output power in real time until the compaction monitoring result meets the standard requirements. Specifically, if the current compaction monitoring value is lower than the target soil compaction characteristic threshold, the power control program extracts the numerical difference between the two as the control deviation, and simultaneously calculates the derivative of this deviation over time as the rate of change of the deviation. The system introduces a fuzzy proportional-integral-derivative (FIG) control algorithm, using the control deviation and its rate of change as input variables for fuzzification processing. Fuzzy inference is then performed using a preset expert fuzzy control rule table, and the dynamic compensation coefficient is output after defuzzification. The mathematical expression for the excitation power control with dynamic compensation coefficients is: ; In the formula, This represents the total amount of control commands output by the system to the exciter at the current moment. This represents the control deviation between the target soil compaction characteristic threshold and the current compaction monitoring value. Represents the proportional control coefficient. Represents the integral control coefficient. Represents the differential control coefficient. This represents the dynamic compensation coefficient calculated after defuzzification using a fuzzy inference engine. The power control program converts the total control command into pulse width modulation duty cycle commands or hydraulic flow servo valve opening control commands, increasing the exciter's drive frequency or eccentric torque amplitude in real time, thereby enhancing the overall output power. During the power increase and compaction process, the piezoelectric sensor continuously performs high-frequency acquisition, and the signal analysis module and power control program synchronously perform high-speed iterative calculations, forming a closed-loop feedback adjustment mechanism until the compaction monitoring results reach and stabilize within the standard requirements. The significant benefit of this step is that the fuzzy proportional-integral-derivative control algorithm effectively overcomes the system hysteresis and overshoot problems caused by the nonlinear changes in soil medium stiffness. In the initial stage of compaction when the soil is relatively loose, the system outputs a smaller power to prevent soil structure damage or excessive compaction; as the soil gradually compacts, the system smoothly and adaptively increases the excitation power, ensuring that energy transfer is always in an optimal matching state. This fully autonomous, flexible, closed-loop control mechanism completely replaces the manual, blind power switching that relies on operator experience. It not only ensures the uniformity of compaction and the quality of the entire work surface, but also minimizes the impact damage and energy waste caused by local over-compaction to the underlying pipelines and mechanical structures.

[0048] Example 6 Further explanation in conjunction with Example 1, such as Figure 1-4 The structure shown, in S4, includes the following steps for adaptively adjusting the spatial angle: The fourth camera 10 continuously captures images of the compaction position at the end of the compaction device 5. The steps are as follows: The fourth camera 10 is configured with a global shutter and automatic exposure control logic to continuously capture high-resolution images of the ground area behind the leveling seat 511 that has just completed the compaction operation at a set frame rate; during the image capture process, the current timestamp of the system and the three-dimensional spatial coordinates of the compaction device 5 are recorded simultaneously to form a continuous image sequence with pose and time label; then, the continuous image sequence with pose and time label is transmitted losslessly through a high-speed industrial communication interface and input into the data buffer of the machine vision inspection module deployed in the industrial control computer. The machine vision detection module uses an image edge feature extraction algorithm to visually detect whether there is unevenness on the ground surface. The steps are as follows: The machine vision detection module extracts a continuous image sequence from the data buffer. First, it uses a Gaussian blur filtering algorithm to smooth the image matrix to remove high-frequency noise interference. Subsequently, the Canny edge detection algorithm was used to calculate the gradient magnitude and gradient direction of the pixel gray values ​​in the image matrix, and non-maximum suppression and double threshold edge connection were performed to extract the set of pixels at the edge of the land surface contour. By combining the Hough linear transform algorithm, the extracted set of pixels on the surface contour edge is mapped to the parameter space, and the physical contour of the actual surface is detected and fitted. Extract the tilt angle and flatness features of the physical contour of the straight line, compare the tilt angle of the physical contour of the straight line with the preset horizontal reference threshold, and if the tilt angle difference exceeds the tolerance range of the horizontal reference threshold, visual detection will detect the unevenness of the ground surface due to slope deviation. If a slope deviation is detected visually, the attitude control module calculates the angle compensation variable and outputs an analog signal to adjust the extension or retraction of the second telescopic cylinder 13, so that the compaction device 5 produces a corresponding angle change to conform to the horizontal reference plane. The steps are as follows: the attitude control module extracts the spatial angle between the actual physical contour of the straight line on the ground and the horizontal reference plane as the initial error value. The PID closed-loop control algorithm is invoked, and the initial error value is input into the proportional-integral-derivative controller. Combined with historical attitude feedback data, the angle compensation variable used to eliminate slope deviation is calculated. According to the inverse kinematics model of the overall suspension structure of the compaction device 5, the angle compensation variable is nonlinearly transformed into the target linear displacement parameter of the second telescopic cylinder 13. The digital-to-analog conversion unit in the attitude control module converts the target linear displacement parameter into a continuous voltage analog signal or current analog signal, which drives the hydraulic proportional valve through a servo amplifier. The hydraulic proportional valve adjusts the flow rate and direction of the hydraulic oil injected into the second telescopic cylinder 13 in real time according to the analog signal, and accurately executes the extension or retraction of the second telescopic cylinder 13. This drives the support frame 505 of the compaction device 5 to rotate around the hinge axis of the lower frame 3, so that the compaction device 5 produces a precise corresponding angle change to dynamically fit the horizontal reference surface, thereby completing the closed-loop automatic correction of the leveling surface.

[0049] In the step of adaptively adjusting the spatial angle, the fourth camera continuously captures images of the compaction position at the end of the compaction device. Specifically, the system configures a global shutter and automatic exposure control logic for the fourth camera to continuously capture high-resolution images of the surface area behind the leveling seat that has just undergone compaction at a set frame rate. The synchronous mapping mathematical expression for image capture is: ; In the formula, Representing the A set of data that is fully synchronized across frames. This represents a matrix of two-dimensional high-resolution images captured by the fourth camera. This represents the three-dimensional spatial pose coordinates of the compaction device recorded by the system at that instant. This represents the system's current high-precision timestamp. During image capture, the system synchronously records pose and time tags according to the above logic, forming a continuous image sequence with pose and time tags. Subsequently, the system transmits this sequence losslessly through a high-speed industrial communication interface and inputs it into the data buffer of the machine vision inspection module deployed in the industrial control computer. The beneficial effect of this step is that the global shutter technology completely eliminates the image jelly effect distortion caused by the vibration compaction process of large machinery, ensuring the absolute authenticity of the physical size characteristics of the image. Synchronously recording three-dimensional coordinates and timestamps allows each frame of two-dimensional image to be accurately anchored in physical space, providing a strict time and space alignment benchmark for subsequent pose compensation and preventing lag and misalignment of the control system.

[0050] After receiving the data, the machine vision inspection module uses an image edge feature extraction algorithm to visually detect whether there are uneven areas on the ground surface. The module extracts a continuous image sequence from the data buffer and first applies a Gaussian blur filter algorithm to smooth the image matrix to remove high-frequency noise. The mathematical expression for Gaussian filtering is: ; In the formula, Represents the Gaussian kernel in two-dimensional pixel coordinates and The weight value at that location, The standard deviation parameter represents the degree of smoothness. Subsequently, the system uses the Canny edge detection algorithm to calculate the gradient magnitude and direction of the pixel grayscale values ​​in the image matrix. The formulas for calculating the gradient magnitude and direction are: ; ; In the formula, Represents the gradient magnitude of a pixel. and These represent the partial derivatives of the pixel in the horizontal and vertical directions, respectively. The direction angle represents the gradient. The system combines non-maximum suppression and double-threshold edge connectivity to extract the set of pixels representing the surface contour edges. Then, the system uses the Hough linear transform algorithm to map the extracted set of surface contour edge pixels to the parameter space, detecting and fitting the actual straight physical contour of the surface. The mathematical model of the Hough transform is... In the formula, This represents the perpendicular distance from the origin of the coordinate system to the fitted line. Represents the angle between the perpendicular line and the horizontal axis. and This represents the coordinates of a point within the edge pixel set. The system uses this algorithm to extract the tilt angle and flatness features of the physical contour of a straight line, comparing the tilt angle of the physical contour with a preset horizontal reference threshold. If the tilt angle difference exceeds the tolerance range of the horizontal reference threshold, visual detection indicates an uneven surface with slope deviation. A significant benefit of this processing step is that the combination of Gaussian filtering and the Canny algorithm effectively filters visual interference caused by dust and gravel at the construction site, accurately revealing the true physical boundaries of the compacted surface. The Hough linear transform possesses strong resistance to local defects, robustly fitting a discrete set of edge pixels, potentially with gaps, into a macroscopic straight line, thus transforming the fuzzy concept of flatness into a geometric tilt angle difference that can be precisely quantified and compared by a computer.

[0051] If a slope deviation is detected visually, the attitude control module calculates the angle compensation variable and outputs an analog signal to adjust the extension or retraction of the second telescopic cylinder, causing the compaction device to produce a corresponding angle change to conform to the horizontal reference plane. In practice, the attitude control module extracts the spatial angle between the actual straight physical contour of the ground surface and the horizontal reference plane as the initial error value. The system calls the PID closed-loop control algorithm, inputting the initial error value into the proportional-integral-derivative controller, and calculates the angle compensation variable used to eliminate the slope deviation by combining it with historical attitude feedback data. The calculation formula for the PID control algorithm is: ; In the formula, The angle compensation variable represents the output at the current moment. This represents the initial error value, i.e., the slope deviation angle. Representative proportional coefficient, Represents the coefficients of the integral process. This represents the coefficients of the differential element. Based on the inverse kinematics model of the overall suspension structure of the compaction device, the system nonlinearly transforms the angle compensation variable into the target linear displacement parameter of the second telescopic cylinder. The geometric model formula for the inverse kinematics is: ; In the formula, This represents the target linear displacement parameter, i.e., the absolute physical length, that the second telescopic cylinder needs to achieve. This represents the fixed physical distance from the hinge point of the lower frame to the hinge point at one end of the second telescopic cylinder. This represents the fixed physical distance from the hinge point of the lower frame to the hinge point of the other end of the second telescopic cylinder on the support frame. This represents the initial mechanical structure angle under no-deviation conditions. The digital-to-analog conversion unit within the attitude control module converts the target linear displacement parameters into continuous analog voltage or current signals, which are then used to drive the hydraulic proportional valve via a servo amplifier. The hydraulic proportional valve adjusts the flow rate and direction of the hydraulic oil injected into the second telescopic cylinder in real time based on the analog signals, precisely executing the extension or retraction of the second telescopic cylinder. This causes the support frame of the compaction device to rotate around the hinge axis of the lower frame, resulting in a precisely corresponding angular change in the compaction device to dynamically conform to the horizontal reference plane, thus completing the closed-loop automatic correction of the leveling surface. The ultimate benefit of this step is that the use of a PID closed-loop algorithm completely eliminates the steady-state error and inertial overshoot generated by the heavy-duty hydraulic system when driving large suspension devices, ensuring smooth and accurate angle correction. The inverse kinematics model directly maps the abstract angle compensation to the millimeter-level mechanical displacement of the telescopic cylinder. Combined with the stepless analog adjustment of the hydraulic proportional valve, this achieves dynamic real-time self-correction of the equipment during continuous compaction. This endows the entire machine with a powerful closed-loop self-healing capability for the working surface, completely replacing the traditional manual stringing and leveling surveying, and greatly improving the flatness and engineering quality of the trench bottom layer.

[0052] Example 7 Further explanation in conjunction with Example 1, such as Figure 1-4 The structure shown illustrates that the data processing and control steps in the method are deployed and executed through an intelligent control system software mounted on an onboard industrial computer. Detailed deployment and execution steps include: The intelligent control system software adopts a modular microservice architecture based on the ROS robot operating system, including hardware driver nodes, environment perception and mapping nodes, path planning nodes, density analysis nodes, and attitude servo control nodes. The steps are as follows: compile and deploy the core communication master node of the ROS robot operating system at the Linux operating system level of the vehicle-mounted industrial control computer; construct independent microservice processes by writing code, configuring the hardware driver node as a topic publisher responsible for parsing sensor-level messages and encapsulating actuator instructions; configuring the environment perception and mapping node as a topic subscriber and publisher that fuses multi-source data and generates a 3D spatial topology map; configuring the path planning node, density analysis node, and attitude servo control node as independent service processing units that execute corresponding core algorithms; and achieving decoupled asynchronous data interaction and distributed parallel computing between the microservice processes through the ROS topic publish-subscribe mechanism, service call mechanism, and action communication mechanism. The vehicle-mounted industrial control computer establishes low-level communication with the first camera 6, the first lidar 7, the third camera 9, the second lidar 12, the piezoelectric sensor 10, and the fourth camera 10 via a field Ethernet bus. The steps are as follows: configure the TCP / IP network protocol stack and UDP multicast protocol in the vehicle-mounted industrial control computer; assign fixed static IP addresses to the first camera 6, the first lidar 7, the third camera 9, the second lidar 12, the piezoelectric sensor 10, and the fourth camera 10 through a gigabit industrial switch in the field Ethernet bus, and establish a Socket long connection socket at the application layer; the hardware driver node reads the raw data stream of each sensor port by calling the corresponding driver API interface and using a multi-threaded polling mechanism, serializes and packages the continuous raw data stream according to the ROS standard sensor message format and synchronizes the global timestamp to complete high-bandwidth, low-latency low-level data acquisition and communication interaction. A closed-loop control flow is formed from data acquisition, edge computing, and hardware action execution. The steps are as follows: the bottom-level sensors continuously acquire data on the trench environment characteristics and the machine's operating status. The hardware driver node publishes the serialized and packaged standard messages to the ROS core communication bus network at high frequency. The environment perception and mapping node, path planning node, density analysis node, and attitude servo control node subscribe to the required feature messages in real time. They utilize the edge computing power of the on-board industrial control computer and combine it with the wireless communication network to call the cloud computing cluster to perform large-scale point cloud registration and model rendering calculations. After the calculation is completed, the optimal planned trajectory, excitation power dynamic compensation coefficient, and attitude angle compensation variables are generated and published to the bottom-level action execution node. The bottom-level action execution node reverse-engineers the received control variables into bottom-level pulse signals or analog drive commands to control the first motor 504, the second motor 506, the vibrator 512, and the second telescopic cylinder 13. These commands are then sent to the servo drivers of each hardware mechanism to execute specific physical actions. The environmental and state changes caused by the actions of the hardware mechanism are captured again by the sensor array and seamlessly enter the next control cycle, thereby constructing a closed-loop control flow for the whole machine with high dynamic response and real-time self-correction.

[0053] In the process of deploying and executing the various data processing and control steps in the method through an intelligent control system software mounted on an onboard industrial control computer, the intelligent control system software adopts a modular microservice architecture based on the ROS robot operating system. Specifically, during deployment, developers compile and deploy the core communication master node of the ROS robot operating system at the Linux operating system level of the onboard industrial control computer. The system constructs multiple independent microservice processes through code writing. The hardware driver node is configured as a topic publisher responsible for parsing sensor-level messages and encapsulating actuator instructions; the environmental perception and mapping node is configured as a topic subscriber and publisher that fuses multi-source data and generates a 3D spatial topology map; and the path planning node, density analysis node, and attitude servo control node are configured as independent service processing units executing their respective core algorithms. The microservice processes achieve decoupled asynchronous data interaction and distributed parallel computing through the ROS topic publication / subscription mechanism, service call mechanism, and action communication mechanism. To quantify the efficiency of this distributed asynchronous communication, the total latency model for data transmission and processing between system nodes is defined as follows: ; In the formula, This represents the total latency for any two microservice nodes in the system to complete a full asynchronous data interaction and processing. This represents the time consumed by the publisher node in serializing the underlying message or computation result into a ROS standard message object. This represents the cross-process routing transmission time of data packets within the ROS core communication bus network. This represents the time it takes for the subscriber node to deserialize and parse the received data packet. This represents the pure computation time for subscriber nodes to call independent service processing units to execute core algorithms. The significant benefit of this deployment step is that the modular microservice architecture completely breaks the tightly coupled monolithic serial structure of traditional industrial control software, allowing core functions such as environmental perception, path planning, and low-level control to operate independently. An anomaly or crash in a single algorithm node will never trigger a global system crash, greatly enhancing the robustness of the software system in complex and highly vibrating construction environments. Simultaneously, distributed parallel computing fully utilizes the low-level computing power of the onboard industrial control computer's multi-core processor, eliminating inter-process blocking and waiting phenomena through asynchronous data interaction mechanisms, ensuring the high-frequency and stable operation of complex navigation and control logic.

[0054] The vehicle-mounted industrial control computer establishes low-level communication with the first camera, first LiDAR, third camera, second LiDAR, piezoelectric sensor, and fourth camera via a field Ethernet bus. During network establishment, the system configures a TCP / IP network protocol stack and a UDP multicast communication protocol within the vehicle-mounted industrial control computer's kernel. Through a gigabit industrial switch within the field Ethernet bus, the system statically assigns fixed IP addresses to the first camera, first LiDAR, third camera, second LiDAR, piezoelectric sensor, and fourth camera, and establishes long-lived socket connections at the operating system's application layer. Hardware driver nodes retrieve the corresponding driver API interfaces for each sensor and read the raw data stream from each sensor's physical port using a multi-threaded polling mechanism. The system serializes and packages the continuous raw data stream according to the ROS standard sensor message format and synchronizes a global timestamp, achieving high-bandwidth and low-latency low-level data acquisition and communication interaction. The data throughput evaluation model calculation formula for the low-level communication network is as follows: ; In the formula, This represents the actual data throughput requirement that the field Ethernet bus must stably process per unit of time. This represents the total physical number of multi-source heterogeneous sensors connected to the system. Representing the The size of the raw message data packet generated by a single sensor sampling, Representing the The actual hardware data sampling refresh rate of each sensor This represents the fixed network communication overhead bytes incurred by the TCP / IP network protocol stack and Socket long-connection sockets in maintaining network heartbeats and packet encapsulation. The beneficial effect of the aforementioned underlying communication establishment steps is that, based on gigabit industrial switches and static IP configurations, the industrial Ethernet architecture provides an absolutely ample physical bandwidth channel for the convergence of massive laser point clouds and high-resolution visual video streams. The Socket long-connection technology, combined with a multi-threaded concurrent polling mechanism, completely avoids the additional TCP three-way handshake delay caused by frequent establishment and disconnection of network connections in traditional protocols, ensuring that every frame of raw signal acquired at high frequency by the sensor array can be captured in real-time and completely by the hardware driver node. Most importantly, the forced synchronization of the underlying hardware's global timestamp during data packaging and serialization fundamentally eliminates the timing disorder problem caused by internal hardware crystal oscillator clock drift in multi-source heterogeneous sensors, providing an absolutely reliable basic data source for subsequent high-precision spatial geometric registration and time synchronization fusion of multi-sensor data.

[0055] Based on the aforementioned software architecture and communication links, the system ultimately forms a seamless closed-loop control flow from data acquisition and edge-cloud computing to hardware action execution. In actual engineering execution, the underlying sensors continuously and uninterruptedly acquire data on trench environmental characteristics and the physical state of the machine's operation. Hardware-driven nodes then frequently publish serialized and packaged standard messages to the ROS core communication bus network. Environmental perception and mapping nodes, path planning nodes, density analysis nodes, and attitude servo control nodes subscribe to the required feature messages in real time with precision. The system utilizes the edge computing power of the onboard industrial control computer to perform local microsecond-level rapid solutions for the highly real-time requirements of the underlying servo closed-loop control. Simultaneously, it combines wireless communication networks to call remote cloud computing clusters, offloading the time-consuming large-scale point cloud registration and 3D model rendering computational load to cloud servers for distributed parallel processing. After computation, the system generates the optimal planned trajectory, dynamic compensation coefficients for excitation power, and attitude angle compensation variables, and publishes these as control commands to the underlying action execution nodes. The underlying action execution nodes inversely calculate the received high-dimensional control variables into drive signals recognizable by the underlying hardware. The mathematical expression of the discrete state-space control model for inverse control solution and physical state execution is as follows: ; ; In the formula, Representative at the Within a discrete control cycle, the final control input matrix calculated in reverse by the underlying action execution node is specifically represented as the pulse signal frequency control quantity sent to the first motor and the second motor, and the analog drive command voltage amplitude of the hydraulic proportional valve sent to the exciter and the second telescopic cylinder. The ideal state reference matrix represents the collaborative computing output of the cloud cluster and the edge industrial control computer. It contains the optimal global planning trajectory coordinates and target attitude angle compensation variables. The representation is through the underlying sensor array in the first The actual volume of the machine accurately captured in each sampling period reflects the pose state matrix; A nonlinear inverse kinematics mapping transformation function representing the mechanical characteristics of a specific hydraulic actuator and transmission gear; The inertial transfer matrix represents the physical state of the system when it is not subject to external forces. This represents the direct driving gain matrix of the control input matrix on the physical state of the organism in space; This represents the completely new physical state that the entire system evolves and is resampled in the next extremely short control cycle after each hardware component receives the drive command and actually performs the physical action. The underlying action execution node directly sends the inversely generated underlying pulse signal or analog drive command to the servo driver of each hardware component to execute the specific physical action. The changes in the external environment and the body's attitude state caused by the hardware component's action are immediately captured again at high speed by the sensor array, seamlessly entering the next closed-loop control cycle. The significant benefit of constructing this complete closed-loop control flow is that the deep computing power collaboration between edge computing and cloud cluster computing completely breaks through the hardware computing power ceiling of a single vehicle-mounted device. It not only ensures the local absolute real-time performance of the underlying millisecond-level mechanical servo control and microsecond-level emergency obstacle avoidance, but also achieves computing power freedom for macroscopic three-dimensional complex topology map construction and global path optimization. By tightly coupling multi-dimensional sensor perception, intelligent algorithm decision-making, and underlying hardware execution through a rigorous discrete state transition mathematical model, the entire single leveling and compaction equipment can continuously and frequently self-correct its state based on the unpredictable terrain undulations and variable soil density feedback at the construction site. This fundamentally and significantly improves the trajectory tracking accuracy, leveling operation quality, and real-time self-correction capability of the machine under harsh and unattended working conditions.

[0056] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A single-unit, mobile leveling and compaction device, characterized in that: The front and sides of the compaction vehicle (1) are equipped with multiple first cameras (6) and first lidar (7); The compaction device (5) is installed on the compaction vehicle (3). The compaction vehicle (3) has a third camera (9) and a fourth camera (10) at the front and rear of the bottom. The third camera (9) and the fourth camera (10) face the compaction position at the end of the compaction device (5). A second laser radar (12) is provided on one side of the third camera (9). The second laser radar (12) is used to scan the material distribution at the bottom of the trench. The compaction vehicle (3) is equipped with a second camera (8) at the front middle position and a fifth camera (11) at the rear top. The compaction vehicle (3) is also equipped with positioning equipment.

2. The single-unit walking leveling and compaction device according to claim 1, characterized in that: The frame structure includes an upper frame (2) and a lower frame (3). The upper frame (2) is a parallelepiped frame structure, and the lower frame (3) is a rectangular frame structure. The lower frame (3) and the upper frame (2) are connected and fixed as an integral frame structure through a hinge axis. The bottom of the frame structure is connected to the walking chassis (9); The compaction device (5) is set on the lower frame (3). One side of the support frame (505) of the compaction device (5) is hinged to one side of the lower frame (3). The other side of the support frame (505) of the compaction device (5) is hinged to one end of the second telescopic cylinder (13). The other end of the lower frame (3) is hinged to the position of the hinge axis of the lower frame (3) and the second upper frame (10). The upper frame (2) of the compaction vehicle (1) is equipped with a hydraulic station (101) on the rear side of the upper frame (2), and a control cabinet (102) is equipped on the rear side of the lower frame (3).

3. The single-unit walking leveling and compaction device according to claim 2, characterized in that: The transverse moving body (503) is slidably connected to the support frame (505). The transverse moving body (503) is provided with a first motor (504) on both sides. The gear at the output end of the first motor (504) meshes with the rack (513) on the side of the support frame (505). A guide telescopic column (501) is provided in the middle of the transverse vehicle body (503). The telescopic rod of the lower part of the guide telescopic column (501) passes through the transverse vehicle body (503) and is connected to the rotary drive device. The drive shaft of the rotary drive device is connected to the connecting gantry (509). The connecting gantry (509) is connected to the leveling seat (511) through multiple shock absorbers (510). A second gear (508) is provided on the rotating end of the rotary drive device, and a second motor (506) is provided on one side of the lower part of the guide telescopic column (501). The first gear (507) on the second motor (506) meshes with the second gear (508). The longitudinal moving car body (501) is also equipped with multiple sixth telescopic cylinders (502). The telescopic end of the sixth telescopic cylinder (502) passes through the transverse moving car body (503) and is connected to the telescopic rod end of the lower end of the guide telescopic column (501). The guide telescopic column (501) is a multi-stage telescopic device composed of multi-stage tubes. A vibrator (512) is provided on the leveling seat (511); The bottom of the leveling seat (511) is also equipped with a piezoelectric sensor (10).

4. The construction method of the single-unit walking leveling and compaction device according to claim 3, characterized in that: The method includes: S1. Collect three-dimensional feature data of the trench environment in front, and generate the automatic walking route of the compaction vehicle (1) by combining the positioning and map information; S2. Scan the distribution of materials at the bottom of the trench and plan the leveling and compaction route of the compaction device (5); S3. Monitor vibration feedback parameters during compaction operations in real time and dynamically adjust excitation power based on compaction assessment results; S4. Perform visual inspection of the flatness of the surface after the operation, and adaptively adjust the spatial angle of the compaction device (5) when the slope deviation is identified.

5. The construction method of the single-unit walking leveling and compaction device according to claim 4, characterized in that: The steps for generating the automatic walking route in S1 include: The first camera (6) and the first lidar (7) are used to synchronously collect trench image data and point cloud data; the multi-sensor extrinsic parameter calibration and point cloud fusion algorithm are used to project and map the pixel color information in the image data to the three-dimensional coordinate system of the point cloud data through the camera intrinsic parameter matrix and extrinsic parameter transformation matrix, so as to give the discrete point cloud texture and color features and fuse them to form three-dimensional view data. The three-dimensional modeling module deployed in the control cabinet (102) is invoked. The three-dimensional modeling module is a spatial mapping program built based on the octree voxel grid algorithm. The three-dimensional modeling module takes the three-dimensional view data as the input source, discretizes the continuous three-dimensional space into a set of three-dimensional grids according to the set voxel resolution, and uses the ray casting method to mark the grid nodes in the idle, occupied and unknown states. Based on this, a three-dimensional spatial topology map containing the geometric structure of the entity in front of the trench and the passage boundary is constructed. The random sampling consensus algorithm is used to calculate the normal vectors and fit the plane of the 3D point set in the 3D spatial topology map. The steps are as follows: three non-collinear points are randomly selected from the 3D point set to calculate the parameters of the initial plane model and the corresponding normal vectors; the orthogonal distance from the remaining points in the 3D point set to the initial plane model is calculated, and points with orthogonal distances less than a set distance threshold are included in the inlier set; the above random selection and calculation process is repeated multiple times, and the optimal plane model with the largest number of inlier sets is retained as the fitted trench bottom area, and the normal vector of the optimal plane model is extracted; based on the fitted trench bottom area, the edge contours on both sides are segmented and extracted, and the morphological skeleton extraction algorithm is applied to calculate the geometric midline coordinate sequence along the longitudinal direction as the trench centerline. At the same time, the vertical elevation difference between the surface grid and the bottom area is calculated along the normal direction of the trench centerline to generate depth distribution data. The algorithm planning module generates an automatic walking route for the compaction vehicle (1) to move autonomously based on the trench location, depth distribution data, self-positioning, and pre-loaded map data. The specific implementation steps are as follows: First, the algorithm planning module obtains the preloaded map data and projects the current self-positioning coordinates of the compaction vehicle (1), the calculated center line of the trench and the depth distribution data into the same global cost map after unifying the coordinate system. The dangerous area weight in the global cost map is updated according to the depth distribution data. Secondly, a safe offset distance that conforms to the wheel track of the compaction vehicle (1) is set on both sides of the center line of the trench, and two smooth reference lines parallel to the trench are generated by combining the B-spline curve fitting algorithm. Next, in the global cost map, starting from its own location coordinates, multiple target waypoints are set along the smooth baseline reference line. The global A-star heuristic search algorithm is used to calculate the global barrier-free path node sequence connecting each target waypoint. The steps are as follows: Initialize the open list and the closed list, add the starting point to the open list and calculate the total value of the starting point. The total value is obtained by adding the actual movement cost from the starting point to the current node and the Manhattan distance heuristic cost from the current node to the target waypoint. Select the node with the lowest total value from the list of nodes to be expanded as the current node. If the current node is the target waypoint, backtrack the parent node to generate a global sequence of accessible path nodes. If the current extended node is not the target waypoint, move it to the closed list, and traverse the neighboring nodes around the current extended node that are not blocked by the global cost map danger zone weight. Calculate the total value of each neighboring node and update the parent node pointing to the current extended node. Add the neighboring nodes to the open list and iterate until the target waypoint is found. Finally, the local dynamic window algorithm is used to sample multiple sets of predicted trajectories in the velocity space based on the kinematic constraint model of the compaction vehicle (1). The steps are as follows: Based on the current linear velocity, angular velocity and maximum acceleration / deceleration capability of the compacted vehicle (1), calculate the dynamic speed window range that can be achieved in the next control cycle; Discretized velocity sampling is performed within a dynamic velocity window, and multiple simulated predicted trajectories within a set time window are generated by combining kinematic model deduction. A comprehensive evaluation function is designed that includes azimuth deviation evaluation, distance from obstacle safety distance evaluation, and current driving speed evaluation. The multiple simulated predicted trajectories are calculated and scored by combining the real-time environmental boundary in the three-dimensional spatial topology map. The optimal predicted trajectory with the highest comprehensive evaluation score is selected and continuously output as motor control commands. The automatic walking route for the compaction vehicle (1) to move autonomously is calculated and generated on both sides of the trench.

6. The construction method of a single-unit walking leveling and compaction device according to claim 4, characterized in that: S2 The planning steps for the leveling and compaction operation route include: The third camera (9) and the second lidar (12) are activated to scan the bottom area of ​​the trench simultaneously. A visual and point cloud registration algorithm based on multi-scale feature fusion is used to extract the edge features of the image obtained by the third camera (9) and the geometric features of the depth point cloud obtained by the second lidar (12). The color and texture information of the image are mapped to the depth point cloud to generate three-dimensional terrain point cloud data containing accurate elevation and boundary information, thereby obtaining the distribution of sand and gravel material and the width features of the trench accurately measured. The path planning software deployed inside the control center reads three-dimensional terrain point cloud data and establishes a local spatial coordinate system at the bottom of the trench. The path planning software uses an inverse distance weighted interpolation algorithm to perform spatial difference calculations on the discrete three-dimensional terrain point cloud data and smoothly reconstruct the continuous elevation surface at the bottom of the trench. By combining the preset target leveling elevation of the trench with the continuous elevation surface, a three-dimensional Boolean subtraction operation is performed to integrally solve the accumulation and pitting volume of sand and gravel in different areas, and the local volume distribution of sand and gravel is accurately calculated. The path planning software aims to level the sand and gravel at the bottom of the trench. Based on the local volume distribution of the sand and gravel, it generates a coordinated movement coordinate sequence of the first motor (504) and the second motor (506), and plans a leveling and compaction route to guide the leveling seat (511). The steps are as follows: Extract the coordinates of peak and trough regions in the volume distribution that exceed the set threshold, mark the peak regions as material collection points and the trough regions as filling points; based on the spatial topological relationship between the material collection points and the filling points, apply the ant colony optimization algorithm to search for the globally optimal material handling and compaction traversal sequence. Combining the linear translation kinematics equation of the lateral moving vehicle body (503) with the rotational kinematics equation of the rotary drive device, the globally optimal sequence searched is reverse-solved into a time and position synchronization sequence of the first motor (504) controlling the lateral movement and the second motor (506) controlling the rotation, generating a cooperative movement coordinate sequence, which ultimately constitutes the leveling and compaction running route.

7. The construction method of the single-unit walking leveling and compaction device according to claim 6, characterized in that: The steps involved in generating a cooperative movement coordinate sequence using path planning software include: A data structure is built in memory to construct a two-dimensional discrete grid map of the bottom of the trench. The grid resolution of the two-dimensional discrete grid map is set to match the physical size of the bottom surface of the leveling seat (511). The three-dimensional coordinates of the sand and gravel distribution obtained by scanning are projected onto the grid nodes of the two-dimensional discrete grid map. The average elevation of the corresponding point cloud in the grid node is extracted and the target leveling elevation is subtracted. The difference is calculated as the height weight of the corresponding grid. Different grids are assigned corresponding height weights to identify the surplus and shortage status. A cover-based full-traversal path search algorithm based on ox-plowing cell decomposition is adopted, with the minimization of height-weighted variance as a constraint, to plan continuous planar traversal nodes in a two-dimensional discrete raster map. The steps are as follows: The two-dimensional discrete raster map is divided into multiple simple rectangular cell regions that do not contain internal obstacles; In each rectangular cell region, a reciprocating cleaning path node is generated; a cost function is introduced to evaluate the height weight gradient between adjacent grids. When the reciprocating cleaning reaches the region of abrupt change in height weight, the algorithm forces priority to connect the peak grid with positive height weight to the valley grid with negative height weight, so that the global height weight variance of the leveling seat (511) is dynamically minimized during the leveling process. In this way, the cleaning path nodes of each rectangular cell region are connected, and a continuous planar traversal node is planned. Combining the lifting stroke limitation of the guide telescopic column (501), a Z-axis depth coordinate is added to the continuous planar traversal nodes, and finally the three-dimensional spatial trajectory coordinates are synthesized and sent to the servo controllers of each drive motor. The steps are as follows: Read the real-time elevation data of the location of the plane traversal node, and calculate the target Z-axis depth coordinate based on the compaction and settlement compression amount preset by the vibrator (512); The target Z-axis depth coordinate is compared with the maximum extension and minimum retraction of the guide telescopic column (501). The coordinates that exceed the lifting stroke limit are physically truncated. The processed Z-axis depth coordinate is matrix-stitched with the X-axis and Y-axis coordinates of the plane traversal node to synthesize a three-dimensional spatial trajectory coordinate containing the three-dimensional spatial position and timestamp. The three-dimensional spatial trajectory coordinate is parsed into pulse control signals through the underlying fieldbus and directly sent to the first motor (504), the second motor (506) and the servo controller of the control unit that controls the extension and retraction of the guide telescopic column (501) to execute high-frequency actions.

8. The construction method of a single-unit walking leveling and compaction device according to claim 4, characterized in that: S3 The steps for dynamically adjusting the excitation power include: When the leveling seat (511) performs compaction work, the piezoelectric sensor (10) collects real-time vibration feedback data of the ground surface at high frequency. The steps are as follows: the piezoelectric ceramic induction array built into the piezoelectric sensor (10) generates micro-charge signals under the alternating load of the vibrator (512) and the ground reaction force; the micro-charge signals are impedance transformed and voltage amplified by the charge amplifier on the hardware acquisition board, and input to the high-frequency analog-to-digital converter for continuous discrete sampling; the high-frequency analog-to-digital converter converts the continuous analog voltage into a digital waveform sequence containing high-precision timestamps and vibration acceleration amplitude according to the set Nyquist sampling frequency, forming real-time vibration feedback data and transmitting it to the signal analysis module in the industrial control computer without delay through the industrial Ethernet bus; The signal analysis module within the industrial control computer filters and performs fast Fourier transform on the real-time vibration feedback data to extract feedback energy characteristics and determine whether the current area's compaction meets the standard. The steps are as follows: First, the signal analysis module uses an adaptive Kalman bandpass filter to denoise the digital waveform sequence, filtering out high-frequency resonance from the mechanical body and low-frequency environmental interference noise. Then, a Hanning window function is added to the denoised time-domain digital waveform sequence to reduce spectral leakage. Subsequently, a fast Fourier transform algorithm is called to convert the time-domain waveform sequence into a frequency-domain energy spectrum. The peak amplitude of the excitation fundamental frequency and the amplitude components of the second and third harmonics are extracted from the frequency-domain energy spectrum. A continuous compaction evaluation algorithm is applied to calculate the ratio of the sum of multiple harmonic amplitudes to the excitation fundamental frequency amplitude as the compaction monitoring value, i.e., the extracted feedback energy characteristics. Finally, the calculated compaction monitoring value is compared differentially with the preset target soil compaction characteristic threshold in the system database to determine whether the current area's compaction meets the standard. When the assessment shows that the compaction does not meet the standard, the power control program calculates the dynamic compensation coefficient and increases the output power of the vibrator (512) in real time until the compaction monitoring result meets the standard requirements. The steps are as follows: If the current compaction monitoring value is lower than the target soil compaction characteristic threshold, the power control program extracts the numerical difference between the two as the control deviation; introduces the fuzzy proportional integral derivative control algorithm, uses the control deviation and the rate of change of the deviation as input variables for fuzzification processing, combines the preset expert fuzzy control rule table for fuzzy reasoning, and outputs the dynamic compensation coefficient after defuzzification; the power control program converts the dynamic compensation coefficient into a pulse width modulation duty cycle command or a hydraulic flow servo valve opening control command, and increases the driving frequency or eccentric torque amplitude of the vibrator (512) in real time, thereby improving the overall output power; during the power increase and compaction operation extension process, the piezoelectric sensor (10) continuously performs high-frequency acquisition action in a loop, and the signal analysis module and the power control program perform high-speed iterative calculation in sync to form a closed-loop feedback adjustment mechanism until the compaction monitoring result reaches and stabilizes within the standard requirements range.

9. The construction method of a single-unit walking leveling and compaction device according to claim 4, characterized in that: S4 The steps for adaptively adjusting the spatial angle include: The fourth camera (10) continuously captures images of the compaction position at the end of the compaction device (5). The steps are as follows: The fourth camera (10) is configured with a global shutter and automatic exposure control logic to continuously capture high-resolution images of the ground area behind the leveling seat (511) where the compaction work has just been completed at a set frame rate. During the image capture process, the current timestamp of the system and the three-dimensional spatial coordinates of the compaction device (5) are recorded simultaneously to form a continuous image sequence with pose and time label. Then, the continuous image sequence with pose and time label is transmitted losslessly through the high-speed industrial communication interface and input into the data buffer of the machine vision inspection module deployed in the industrial control computer. The machine vision detection module uses an image edge feature extraction algorithm to visually detect whether there is unevenness on the ground surface. The steps are as follows: The machine vision detection module extracts a continuous image sequence from the data buffer. First, it uses a Gaussian blur filtering algorithm to smooth the image matrix to remove high-frequency noise interference. Subsequently, the Canny edge detection algorithm was used to calculate the gradient magnitude and gradient direction of the pixel gray values ​​in the image matrix, and non-maximum suppression and double threshold edge connection were performed to extract the set of pixels at the edge of the land surface contour. By combining the Hough linear transform algorithm, the extracted set of pixels on the surface contour edge is mapped to the parameter space, and the physical contour of the actual surface is detected and fitted. Extract the tilt angle and flatness features of the physical contour of the straight line, compare the tilt angle of the physical contour of the straight line with the preset horizontal reference threshold, and if the tilt angle difference exceeds the tolerance range of the horizontal reference threshold, visual detection will detect the unevenness of the ground surface due to slope deviation. If the slope deviation is detected by vision, the attitude control module calculates the angle compensation variable and outputs an analog signal to adjust the extension or shortening of the second telescopic cylinder (13) so that the compaction device (5) produces a corresponding angle change to fit the horizontal reference plane. The steps are: the attitude control module extracts the spatial angle between the actual physical outline of the ground surface straight line and the horizontal reference plane as the initial error value. The PID closed-loop control algorithm is called, and the initial error value is input into the proportional integral derivative controller. The angle compensation variable used to eliminate the slope deviation is calculated by combining the historical data of attitude feedback. According to the kinematic inverse solution model of the overall suspension structure of the compaction device (5), the angle compensation variable is nonlinearly converted into the target linear displacement parameter of the second telescopic cylinder (13). The digital-to-analog conversion unit in the attitude control module converts the target linear displacement parameter into a continuous voltage analog signal or current analog signal, and drives the hydraulic proportional valve through the servo amplifier. The hydraulic proportional valve adjusts the flow rate and direction of the hydraulic oil injected into the second telescopic cylinder (13) in real time according to the analog signal, and accurately executes the extension or retraction of the second telescopic cylinder (13), which drives the support frame (505) of the compaction device (5) to rotate around the hinge axis of the lower frame (3), so that the compaction device (5) produces a precise corresponding angle change to dynamically fit the horizontal reference surface, thereby completing the closed-loop automatic correction of the leveling work surface.

10. The construction method of a single-unit walking leveling and compaction device according to claim 4, characterized in that: The various data processing and control steps in the method are deployed and executed through an intelligent control system software mounted on an onboard industrial computer. The detailed deployment and execution steps include: The intelligent control system software adopts a modular microservice architecture based on the ROS robot operating system, including hardware driver nodes, environment perception and mapping nodes, path planning nodes, density analysis nodes, and attitude servo control nodes. The steps are as follows: compile and deploy the core communication master node of the ROS robot operating system at the Linux operating system level of the vehicle-mounted industrial control computer; construct independent microservice processes by writing code, configuring the hardware driver node as a topic publisher responsible for parsing sensor-level messages and encapsulating actuator instructions; configuring the environment perception and mapping node as a topic subscriber and publisher that fuses multi-source data and generates a 3D spatial topology map; configuring the path planning node, density analysis node, and attitude servo control node as independent service processing units that execute corresponding core algorithms; and achieving decoupled asynchronous data interaction and distributed parallel computing between the microservice processes through the ROS topic publish-subscribe mechanism, service call mechanism, and action communication mechanism. The vehicle-mounted industrial control computer establishes low-level communication with the first camera (6), the first lidar (7), the third camera (9), the second lidar (12), the piezoelectric sensor (10), and the fourth camera (10) through the field Ethernet bus. The steps are as follows: configure the TCP / IP network protocol stack and UDP multicast protocol in the vehicle-mounted industrial control computer; assign fixed static IP addresses to the first camera (6), the first lidar (7), the third camera (9), the second lidar (12), the piezoelectric sensor (10), and the fourth camera (10) through the gigabit industrial switch in the field Ethernet bus, and establish a Socket long connection socket in the application layer; the hardware driver node reads the raw data stream of each sensor port by calling the corresponding driver API interface and using a multi-threaded polling mechanism, serializes and packages the continuous raw data stream according to the ROS standard sensor message format and synchronizes the global timestamp to complete the high-bandwidth, low-latency low-level data acquisition and communication interaction. A closed-loop control flow is formed, from data acquisition and edge computing to hardware action execution. The steps are as follows: the bottom-level sensors continuously acquire data on the trench environment characteristics and the machine's operating status; the hardware-driven nodes publish serialized and packaged standard messages at high frequency to the ROS core communication bus network; the environmental perception and mapping nodes, path planning nodes, density analysis nodes, and attitude servo control nodes subscribe to the required feature messages in real time, utilize the edge computing power of the on-board industrial control computer, and combine the wireless communication network to call the cloud computing cluster to perform large-scale point cloud registration and model rendering calculations; after the calculation is completed, the optimal planned trajectory is generated. The dynamic compensation coefficients of the trace and excitation power, as well as the attitude angle compensation variables, are published to the underlying action execution node. The underlying action execution node reverse-calculates the received control variables into underlying pulse signals or analog drive commands to control the first motor (504), the second motor (506), the exciter (512), and the second telescopic cylinder (13), and sends them to the servo drivers of each hardware mechanism to execute specific physical actions. The environmental and state changes caused by the actions of the hardware mechanism are captured again by the sensor array and seamlessly enter the next control cycle, thereby constructing a closed-loop control flow of the whole machine with high dynamic response and real-time self-correction.

Images