Multi-terrain-based photovoltaic cleaning crawler stability adaptive control method and system

Abstract

The invention provides a stability self-adaptive control method and system based on a multi-terrain photovoltaic cleaning crawler, and relates to the technical field of photovoltaic cleaning robots, and the method comprises the steps: obtaining multi-source sensing data of the crawler, including attitude data, crawler pressure data, cleaning arm stress data and terrain scanning data; constructing a terrain three-dimensional point cloud model and carrying out adaptive grid division; identifying terrain types and feature parameters based on multi-feature fusion, and calculating a stability index of the crawler in combination with multi-source data; a self-adaptive entropy enhancement cooperative tracking algorithm is adopted to solve crawler torque distribution parameters and cleaning arm pose parameters; the method comprises the following steps: constructing a dynamic space-time constraint cost function, and obtaining an optimal path sequence by adopting progressive rolling time domain planning; and a crawler advancing instruction and a cleaning arm posture instruction are generated in combination with the prediction control strategy, and stable operation control over the photovoltaic cleaning crawler is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic cleaning robots, and in particular to a stability adaptive control method and system for a multi-terrain photovoltaic cleaning crawler vehicle. Background Art

[0002] Photovoltaic power generation, a crucial component of clean energy, directly impacts its efficiency by the surface cleanliness of its panels. Currently, large-scale photovoltaic power plants commonly utilize tracked cleaning vehicles for panel cleaning and maintenance. Since photovoltaic power plants are often located in complex terrain, such as mountainous areas and deserts, cleaning operations face stability challenges in these complex environments. To ensure cleaning quality and equipment safety, tracked cleaning vehicles require precise stability control in diverse terrains.

[0003] The existing control methods for photovoltaic cleaning crawlers have the following main shortcomings: traditional control methods are mostly based on single sensor data, lack the ability to integrate and process multi-source information, and have difficulty in accurately perceiving complex terrain features; existing stability control strategies often use fixed parameter configurations and cannot be adaptively adjusted to different terrain features; path planning methods generally ignore the impact of terrain changes on the dynamic characteristics of crawlers, resulting in the risk of instability during actual operations; the cleaning arm posture control and crawler motion control are relatively independent, and lack a collaborative optimization mechanism.

[0004] In summary, to address the stability control issues of photovoltaic tracked cleaning vehicles in complex terrain environments, this paper establishes a terrain perception model that integrates multi-source sensor information, combines it with an adaptive entropy-enhanced collaborative tracking algorithm to optimize track torque distribution and cleaning arm posture parameters, and employs a progressive path planning method with dynamic spatiotemporal constraints to achieve stable operational control of the tracked vehicle in complex terrain environments. This invention effectively improves the environmental adaptability and operational safety of cleaning equipment and has important engineering application value. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for adaptively controlling the stability of a multi-terrain photovoltaic cleaning crawler vehicle, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for adaptive stability control of a multi-terrain photovoltaic cleaning crawler vehicle is provided, comprising:

[0008] Obtain the photovoltaic cleaning crawler vehicle's posture data, crawler pressure data, cleaning arm force data, and terrain scanning data to construct a three-dimensional point cloud model of the terrain;

[0009] Adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain feature parameters based on multi-feature fusion;

[0010] Calculate the crawler vehicle stability index based on posture data, crawler pressure data and cleaning arm force data;

[0011] According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, an adaptive entropy enhanced collaborative tracking algorithm is implemented to solve the track torque distribution parameters and cleaning arm posture parameters.

[0012] A dynamic spatiotemporal constraint cost function is constructed based on track torque distribution parameters and terrain characteristic parameters, and the optimal path sequence is obtained through progressive rolling horizon planning.

[0013] According to the optimal path sequence and cleaning arm posture parameters, the crawler vehicle travel instructions and cleaning arm posture instructions are generated in combination with the predictive control strategy to achieve stable operation of the photovoltaic cleaning crawler vehicle.

[0014] In an optional embodiment,

[0015] Adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain characteristic parameters based on multi-feature fusion, including:

[0016] Project the terrain 3D point cloud model onto the horizontal plane, calculate the initial side length based on the total area and total number of point clouds, and divide it into multiple initial grid units;

[0017] Calculate the point cloud density distribution of the initial grid cell, construct the density flow field based on the density gradient, and calculate the deformation vector to generate the deformed grid. When the strain energy density of the deformed grid exceeds the preset density threshold, split along the main direction of the density flow field to generate the target grid cell;

[0018] Calculate the height range within the target grid cell to determine the height difference value, calculate the height change rate in the orthogonal direction to determine the height gradient; establish the height correlation characteristics between target grid cells, calculate the height difference coefficient of adjacent cells, and construct a spatial correlation matrix;

[0019] The height difference value, height gradient and spatial correlation matrix are combined into a feature vector, which is input into the multi-feature fusion terrain classification discriminant function to determine the terrain type of the target grid cell based on the conditional probability maximization criterion;

[0020] The terrain slope of the target grid cell is calculated based on the height gradient, and the terrain roughness is determined by calculating the integral deviation between the actual height and the least squares fitting plane;

[0021] The terrain type, terrain slope and terrain roughness are combined into a terrain characteristic parameter vector to determine the terrain characteristic parameters of the target grid cell.

[0022] In an optional embodiment,

[0023] Calculate the point cloud density distribution of the initial grid cell, construct the density flow field based on the density gradient, and calculate the deformation vector to generate the deformed grid. When the strain energy density of the deformed grid exceeds the preset density threshold, it is split along the main direction of the density flow field to generate the target grid cells including:

[0024] The number of point clouds in the initial grid cells is counted, and the number of point clouds per unit area is calculated to obtain the point cloud density distribution. The density change rate in the X direction and the density change rate in the Y direction are calculated according to the point cloud density of adjacent initial grid cells to form a density gradient vector.

[0025] Perform an orthogonal rotation on the density gradient vector to obtain an orthogonal rotation vector, perform a linear combination of the density gradient vector and the orthogonal rotation vector to construct a density flow field, and calculate the grid deformation vector according to the direction of the density flow field and the modulus of the density gradient vector;

[0026] Apply the mesh deformation vector to the vertices of the initial mesh unit to generate a deformed mesh unit, and calculate the deformation work per unit area in the deformed mesh unit to obtain the strain energy density;

[0027] When the strain energy density exceeds a preset density threshold, the deformed grid unit is split along the main direction of the density flow field, the difference between the strain energy density and the preset density threshold is calculated, and the difference is substituted into an exponential function to determine the size of the split subgrid;

[0028] The mesh side length ratio, mesh area change rate, and deformation vector difference of adjacent mesh vertices of the split sub-mesh are calculated. When the mesh side length ratio does not exceed the preset distortion threshold, the mesh area change rate does not exceed the preset area change threshold, and the deformation vector difference does not exceed the product of the preset smoothing coefficient and the mesh vertex distance, the target mesh unit is determined.

[0029] In an optional embodiment,

[0030] The crawler vehicle stability index is calculated based on posture data, track pressure data and cleaning arm force data, including:

[0031] Receive attitude data, including pitch angle, roll angle and heading angle; receive track pressure data, including left track pressure distribution data and right track pressure distribution data; receive cleaning arm force data, including torque data and joint axial force data of each joint of the cleaning arm;

[0032] Determining the coordinates of the boundary points of the crawler vehicle support polygon based on the posture data, calculating the distances from the projection point of the crawler vehicle's center of gravity to the sides of the crawler vehicle support polygon, and determining the ratio of the minimum value of the distances to the characteristic size of the crawler vehicle support polygon as the static stability margin;

[0033] Calculating the zero moment point coordinates of the crawler vehicle based on the track pressure data and the cleaning arm force data, and determining the dynamic stability margin as a complementary value of the ratio of the distance between the zero moment point coordinates and the center coordinates of the crawler support polygon to a preset safety radius;

[0034] The weighted sum of the static stability margin and the dynamic stability margin is determined as the tracked vehicle stability index.

[0035] In an optional embodiment,

[0036] According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, the adaptive entropy enhanced collaborative tracking algorithm is executed to solve the track torque distribution parameters and cleaning arm posture parameters, including:

[0037] Based on a preset track torque distribution parameter range and a preset cleaning arm posture parameter range, a parameter search space is determined, the parameter search space is divided into multiple equally probable subspaces, and a collaborative tracking agent is generated in each equally probable subspace based on uniformly distributed random sampling. All the collaborative tracking agents generated in the equally probable subspaces constitute a collaborative tracking agent group;

[0038] Calculating a local entropy value for each collaborative tracking agent in the collaborative tracking agent group based on the neighborhood probability density, and determining the mean of the local entropy values as the overall entropy value of the group;

[0039] According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, a fitness function is constructed by weighted combination of terrain evaluation value and stability evaluation value;

[0040] Calculating an adaptive step-size adjustment coefficient based on the overall entropy value of the group, selecting a first agent according to the fitness function, and updating the parameters of the agents in the collaborative tracking agent group in combination with the adaptive step-size adjustment coefficient;

[0041] Calculating the parameter dispersion of the collaborative tracking agent group, determining a group diversity index, and performing Gaussian random mutation when the group diversity index is less than a preset diversity threshold;

[0042] When the change in the optimal parameters of continuous iterations is less than the preset convergence threshold, the current optimal crawler torque distribution parameters and cleaning arm posture parameters are output;

[0043] The weight coefficient of the fitness function is dynamically adjusted based on the terrain type and terrain characteristic parameters updated in real time.

[0044] In an optional embodiment,

[0045] Based on the track torque distribution parameters and terrain characteristic parameters, a dynamic spatiotemporal constraint cost function is constructed. The optimal path sequence is obtained through progressive hierarchical prediction time domain planning, including:

[0046] Based on the track torque distribution parameters, the torque distribution cost term is calculated through the left track torque parameters and the right track torque parameters. Based on the terrain characteristic parameters, the terrain characteristic cost term is calculated through the terrain slope parameter and the terrain roughness parameter. The dynamic spatiotemporal constraint cost function is constructed by combining the time cost term and the path smoothness cost term.

[0047] Within the overall planning horizon, a hierarchical prediction horizon architecture is constructed, the prediction horizon length is adaptively adjusted, and the local optimal control sequence that minimizes the dynamic spatiotemporal constraint cost function is solved. The corresponding control instructions are executed and the prediction horizon is updated in a sliding manner. The loop is iterated until the overall planning horizon ends and the planning result is obtained.

[0048] Based on the planning results, an optimal path sequence consisting of position parameters, heading angle parameters, linear velocity parameters, and angular velocity parameters is generated.

[0049] In an optional embodiment,

[0050] Within the overall planning horizon, a hierarchical prediction horizon architecture is constructed, the prediction horizon length is adaptively adjusted, and the local optimal control sequence that minimizes the dynamic spatiotemporal constraint cost function is solved. The corresponding control instructions are executed and the prediction horizon is updated in a sliding manner. The loop is iterated until the overall planning horizon ends. The planning results include:

[0051] The overall planning time domain is divided into a macro-planning layer time domain and a micro-execution layer time domain, wherein the macro-planning layer time domain includes multiple consecutive prediction cycles, and each prediction cycle corresponds to a micro-execution layer time domain;

[0052] Collecting environmental information of the current location to obtain environmental complexity, detecting a deviation between the actual trajectory and the planned trajectory, determining a length of a predicted time domain based on the environmental complexity, and determining an adjustment interval for the predicted time domain based on the deviation;

[0053] Dividing the current prediction time domain into two adjacent time intervals according to a preset ratio, determining a first time interval and a second time interval, establishing a dynamic prediction model including track-ground interaction forces to generate a motion trajectory in the first time interval, establishing a kinematic prediction model based on steering constraints to generate a path direction in the second time interval, and combining the motion trajectory and the path direction to generate a predicted trajectory;

[0054] Based on the predicted trajectory, a control sequence is solved to minimize the dynamic spatiotemporal constraint cost function, the control instruction corresponding to the first time interval is executed, the predicted time domain is updated to the starting point of the next prediction, and the iterative solution is continued until the total planning time domain ends to obtain the planning result.

[0055] According to a second aspect of the embodiments of the present invention,

[0056] Provided is a stability adaptive control system for a multi-terrain photovoltaic cleaning crawler vehicle, comprising:

[0057] The first unit is used to obtain the posture data, track pressure data, cleaning arm force data and terrain scanning data of the photovoltaic cleaning crawler vehicle to construct a three-dimensional point cloud model of the terrain;

[0058] The second unit is used to adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain feature parameters based on multi-feature fusion;

[0059] The third unit is used to calculate the stability index of the crawler vehicle based on the posture data, the crawler pressure data and the cleaning arm force data;

[0060] The fourth unit is used to execute the adaptive entropy enhanced collaborative tracking algorithm to solve the track torque distribution parameters and the cleaning arm posture parameters according to the terrain type, terrain characteristic parameters and the tracked vehicle stability index;

[0061] The fifth unit is used to construct a dynamic spatiotemporal constraint cost function based on the track torque distribution parameters and terrain characteristic parameters, and obtain the optimal path sequence through progressive rolling horizon planning;

[0062] The sixth unit is used to generate crawler vehicle travel instructions and cleaning arm posture instructions based on the optimal path sequence and cleaning arm posture parameters in combination with the predictive control strategy to achieve stable operation of the photovoltaic cleaning crawler vehicle.

[0063] According to a third aspect of the embodiments of the present invention,

[0064] An electronic device is provided, comprising:

[0065] processor;

[0066] a memory for storing processor-executable instructions;

[0067] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0068] According to a fourth aspect of the embodiments of the present invention,

[0069] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0070] In an embodiment of the present invention, by constructing a three-dimensional point cloud model of the terrain and adaptive meshing technology, different terrain types and characteristic parameters can be accurately identified, and accurate perception and analysis of complex and changeable terrain can be achieved, thereby improving the adaptability of the photovoltaic cleaning crawler vehicle in different photovoltaic scenarios, enabling it to cope with various complex terrain environments such as slopes and uneven roads; the stability index is calculated by combining posture data, track pressure data and cleaning arm force data, and the track torque distribution and cleaning arm posture are dynamically adjusted through an adaptive entropy enhanced collaborative tracking algorithm, thereby realizing real-time monitoring and active adjustment of the track vehicle stability, effectively preventing the risk of rollover, and improving the safety and reliability of the equipment under extreme working conditions; the dynamic space-time constraint cost function and the progressive rolling time domain planning method are used to generate the optimal path sequence, and the operating parameters are adjusted in real time in combination with the predictive control strategy, which not only optimizes the operating efficiency, but also realizes energy consumption optimization and wear reduction during the cleaning process, extends the equipment service life, reduces maintenance costs, and at the same time improves the cleaning quality of the photovoltaic panels and the overall efficiency of the cleaning operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of a flow chart of a method for adaptively controlling stability of a multi-terrain photovoltaic cleaning crawler vehicle according to an embodiment of the present invention;

[0072] Figure 2 is the relationship diagram between mesh strain energy density and deformation performance;

[0073] Figure 3 This is a comparison chart of tracked vehicle terrain adaptability results;

[0074] Figure 4 This is the effect diagram of adaptive temporal adjustment during complex terrain navigation. DETAILED DESCRIPTION

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0076] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0077] Figure 1 FIG. 1 is a flow chart of a method for adaptively controlling the stability of a multi-terrain photovoltaic cleaning crawler vehicle according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0078] Obtain the photovoltaic cleaning crawler vehicle's posture data, crawler pressure data, cleaning arm force data, and terrain scanning data to construct a three-dimensional point cloud model of the terrain;

[0079] Adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain feature parameters based on multi-feature fusion;

[0080] Calculate the crawler vehicle stability index based on posture data, crawler pressure data and cleaning arm force data;

[0081] According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, an adaptive entropy enhanced collaborative tracking algorithm is implemented to solve the track torque distribution parameters and cleaning arm posture parameters.

[0082] A dynamic spatiotemporal constraint cost function is constructed based on track torque distribution parameters and terrain characteristic parameters, and the optimal path sequence is obtained through progressive rolling horizon planning.

[0083] According to the optimal path sequence and cleaning arm posture parameters, the crawler vehicle travel instructions and cleaning arm posture instructions are generated in combination with the predictive control strategy to achieve stable operation of the photovoltaic cleaning crawler vehicle.

[0084] In an optional embodiment, adaptively meshing the terrain three-dimensional point cloud model to obtain multiple grid cells, establishing highly correlated features between the grid cells, and determining the terrain type and terrain feature parameters based on multi-feature fusion includes:

[0085] Project the terrain 3D point cloud model onto the horizontal plane, calculate the initial side length based on the total area and total number of point clouds, and divide it into multiple initial grid units;

[0086] Calculate the point cloud density distribution of the initial grid cell, construct the density flow field based on the density gradient, and calculate the deformation vector to generate the deformed grid. When the strain energy density of the deformed grid exceeds the preset density threshold, split along the main direction of the density flow field to generate the target grid cell;

[0087] Calculate the height range within the target grid cell to determine the height difference value, calculate the height change rate in the orthogonal direction to determine the height gradient; establish the height correlation characteristics between target grid cells, calculate the height difference coefficient of adjacent cells, and construct a spatial correlation matrix;

[0088] The height difference value, height gradient and spatial correlation matrix are combined into a feature vector, which is input into the multi-feature fusion terrain classification discriminant function to determine the terrain type of the target grid cell based on the conditional probability maximization criterion;

[0089] The terrain slope of the target grid cell is calculated based on the height gradient, and the terrain roughness is determined by calculating the integral deviation between the actual height and the least squares fitting plane;

[0090] The terrain type, terrain slope and terrain roughness are combined into a terrain characteristic parameter vector to determine the terrain characteristic parameters of the target grid cell.

[0091] In one specific embodiment, the acquired three-dimensional point cloud model of the terrain is projected onto a horizontal plane, and the initial grid side length is calculated based on the total area and total number of points in the point cloud. Assuming that a terrain point cloud contains 100,000 points and covers an area of 1 square kilometer, the initial grid side length can be set to 10 meters, thereby dividing the entire area into 10,000 initial grid cells. The formula for calculating the initial grid side length is: the initial side length is equal to the square root of the area covered by the point cloud divided by the grid division coefficient. The division coefficient can be adjusted according to actual application requirements and is typically set to 10 to 20.

[0092] For each initial grid cell, calculate its point cloud density distribution. For example, for the grid cell numbered (25,30), the number of internal points is 124, and the area is 100 square meters, so its point cloud density is 1.24 points / square meter. The adjacent grid cell numbered (25,31) has 78 points, with a density of 0.78 points / square meter. Based on the density values of the grid cells, construct a density distribution map of the entire area and calculate the density gradient. For adjacent grid cells, the density gradient is the density difference divided by the distance between the grid centers. For example, the density gradient of the two grid cells mentioned above is (1.24-0.78) / 10=0.046 / square meter / meter.

[0093] The density flow field is constructed using density gradient information, and the deformation vector of each grid cell is calculated based on it. The deformation vector is consistent with the direction of the density gradient, and its size is proportional to the density gradient. The proportional coefficient can be set to 0.5. The initial grid is deformed according to the deformation vector to obtain a deformed grid. The strain energy density of the deformed grid is calculated. When the strain energy density exceeds a preset threshold (such as 0.15), the grid is split along the main direction of the density flow field. For example, the strain energy density of the grid cell near the ridge line is 0.23, which exceeds the threshold, and it is split into two sub-grids. After splitting, the final grid is called the target grid cell, and the total number is usually 1.2 to 1.5 times the number of initial grids.

[0094] For each target grid cell, calculate the height characteristics of the internal point cloud and the height range, which is the difference between the highest and lowest points. For example, if the highest point in a target grid cell is 325.8 meters and the lowest point is 318.2 meters, the height range is 7.6 meters. This value is saved as the height difference value.

[0095] Calculate the rate of change of height in the orthogonal directions within the grid cell to determine the height gradient. Select the east-west and north-south directions within the grid cell as orthogonal directions and calculate the rate of change of height in the two directions. For example, the heights of the two end points in the east-west direction are 321.5 meters and 324.3 meters respectively, and the grid length is 12 meters. The rate of change of height in the east-west direction is (324.3-321.5) / 12=0.23 meters / meter; similarly, the rate of change of height in the north-south direction is calculated to be 0.31 meters / meter. Take the average of the two, 0.27 meters / meter, as the height gradient of the grid cell.

[0096] Establish height correlation features between target grid cells and calculate the height difference coefficient between adjacent cells. For two adjacent grid cells, the height difference coefficient is the absolute value of their average height difference divided by the average of their height range differences. For example, if the average heights of grid A and grid B are 322.4 meters and 320.1 meters, respectively, and the height ranges are 7.6 meters and 6.8 meters, respectively, their height difference coefficient is |322.4 - 320.1| / ((7.6 + 6.8) / 2) = 2.3 / 7.2 = 0.32.

[0097] A spatial correlation matrix is constructed based on the spatial positional relationships between target grid cells. For example, for a 5×5 grid area, the spatial correlation between the central grid cell and the surrounding 24 grid cells is divided into three categories based on their distance: the correlation between the eight adjacent grid cells is 0.8, the correlation between the grid cells one grid away is 0.5, and the correlation between the remaining grid cells is 0.2.

[0098] The height difference value, height gradient, and spatial correlation matrix of each target grid cell are combined into a feature vector and input into a multi-feature fusion terrain classification discriminant function. This function, based on a Bayesian conditional probability model, calculates the conditional probability that the feature vector belongs to each terrain type and selects the type with the highest probability as the terrain type for that grid cell. For example, if the conditional probabilities for the four terrain types, namely, flatland, mountainous, hilly, and gully, corresponding to the feature vector of a grid cell, are 0.15, 0.62, 0.18, and 0.05, respectively, the terrain type of that grid cell is classified as mountainous.

[0099] Calculates the terrain slope of the target grid cell based on the altitude gradient. The terrain slope is equal to the inverse tangent of the altitude gradient, expressed as an angle. For example, if the altitude gradient is 0.27 m / m, the slope is 15.1 degrees.

[0100] Calculate the terrain roughness of the target grid cell. First, use the least squares method to fit the plane equation of the point cloud within the grid cell. Then calculate the deviation between the actual height of each point and the height of the fitted plane. Take the root mean square value of these deviations as the roughness index. For example, if the sum of the squares of the height deviations of 10 points in a grid cell is 1.82 square meters, the roughness is (1.82 / 10). 1 / 2 =0.43 meters.

[0101] Terrain type, slope, and roughness are combined into a terrain characteristic parameter vector, which is used as the terrain characteristic parameter of the target grid cell. For example, the terrain characteristic parameter vector of a grid cell is {Terrain type: mountainous, slope: 15.1 degrees, roughness: 0.43 meters}.

[0102] Adaptive grid division driven by density flow field can better preserve the characteristic details of the terrain, and combined with the classification and discrimination method of multi-feature fusion, it can achieve accurate identification and parameter extraction of complex terrain, providing key technical support for applications such as drone route planning and autonomous driving terrain adaptability assessment.

[0103] In an optional embodiment, calculating the point cloud density distribution of the initial grid cell, constructing a density flow field based on the density gradient, and calculating the deformation vector to generate a deformed grid, and when the strain energy density of the deformed grid exceeds a preset density threshold, splitting along the main direction of the density flow field to generate a target grid cell includes:

[0104] The number of point clouds in the initial grid cells is counted, and the number of point clouds per unit area is calculated to obtain the point cloud density distribution. The density change rate in the X direction and the density change rate in the Y direction are calculated according to the point cloud density of adjacent initial grid cells to form a density gradient vector.

[0105] Perform an orthogonal rotation on the density gradient vector to obtain an orthogonal rotation vector, perform a linear combination of the density gradient vector and the orthogonal rotation vector to construct a density flow field, and calculate the grid deformation vector according to the direction of the density flow field and the modulus of the density gradient vector;

[0106] Apply the mesh deformation vector to the vertices of the initial mesh unit to generate a deformed mesh unit, and calculate the deformation work per unit area in the deformed mesh unit to obtain the strain energy density;

[0107] When the strain energy density exceeds a preset density threshold, the deformed grid unit is split along the main direction of the density flow field, the difference between the strain energy density and the preset density threshold is calculated, and the difference is substituted into an exponential function to determine the size of the split subgrid;

[0108] The mesh side length ratio, mesh area change rate, and deformation vector difference of adjacent mesh vertices of the split sub-mesh are calculated. When the mesh side length ratio does not exceed the preset distortion threshold, the mesh area change rate does not exceed the preset area change threshold, and the deformation vector difference does not exceed the product of the preset smoothing coefficient and the mesh vertex distance, the target mesh unit is determined.

[0109] In a specific embodiment, a point cloud density analysis is performed on the initial grid unit, the entire point cloud data is projected onto the horizontal XY plane, and divided into a number of initial grid units according to the preset initial grid side length. For example, for point cloud data covering an area of 5,000 square meters, an initial grid side length of 10 meters is selected, and the entire area is divided into 50 initial grid units. The number of point clouds contained in each initial grid unit is counted and divided by the area of the grid unit to obtain the point cloud density distribution. For example, if the area of the grid unit numbered (3,4) is 100 square meters and it contains 225 point clouds, the point cloud density of the grid unit is 2.25 points per square meter.

[0110] For each initial grid cell, calculate the density change rate in the X and Y directions. For example, for grid cell (3,4), the point cloud density of its adjacent grid cell (4,4) in the X direction is 1.85 points / m2. Therefore, the density change rate in the X direction is (1.85-2.25) / 10 = -0.04 points / m2 / m. Similarly, the point cloud density of the adjacent grid cell (3,5) in the Y direction is 2.75 points / m2. Therefore, the density change rate in the Y direction is (2.75-2.25) / 10 = 0.05 points / m2 / m. The density change rates in the X and Y directions together form the density gradient vector for this grid cell: (-0.04, 0.05) points / m2 / m.

[0111] Perform an orthogonal rotation on the density gradient vector to obtain an orthogonal rotation vector. An orthogonal rotation rotates the density gradient vector 90 degrees counterclockwise, rotating the vector (dx, dy) to (-dy, dx). For the density gradient vector (-0.04, 0.05) in the above grid cell, its orthogonal rotation vector is (-0.05, -0.04) per square meter per meter.

[0112] The density flow field is constructed by linearly combining the density gradient vector with the orthogonal rotation vector. The coefficients of the linear combination are α and β, respectively. α is typically 0.7 and β is 0.3. For the above grid cell, the density flow field vector is 0.7 × (-0.04, 0.05) + 0.3 × (-0.05, -0.04) = (-0.043, 0.023) per square meter per meter.

[0113] The grid deformation vector is calculated based on the direction of the density flow field and the modulus of the density gradient vector. The modulus of the density gradient vector is ((-0.04) 2 +(0.05) 2 ) 1 / 2 = 0.064 / m2 / m. The size of the deformation vector is proportional to the density gradient modulus, and its direction is consistent with the density flow vector. The proportional coefficient γ is usually set to 8. For the above grid unit, the grid deformation vector is 8×0.064×(-0.043,0.023) / ((-0.043) 2 +(0.023) 2 ) 1 / 2 =(-0.333,0.179) meters.

[0114] Apply the mesh deformation vector to each vertex of the initial mesh cell to generate a deformed mesh cell. For example, if the coordinates of the four vertices of the initial mesh cell are (30, 40), (40, 40), (40, 50), and (30, 50), the deformed coordinates are (29.667, 40.179), (39.667, 40.179), (39.667, 50.179), and (29.667, 50.179).

[0115] Calculate the strain energy density of the deformed mesh element. Strain energy density is equal to the deformation work per unit area and can be calculated from the vertex displacement before and after mesh deformation and the corresponding density gradient force. For the mesh element above, assume the calculated strain energy density is 0.085.

[0116] Determine whether the strain energy density exceeds a preset density threshold, typically set at 0.08. If so, split the deformed mesh element along the principal direction of the density flow field. For the mesh element described above, 0.085 is greater than 0.08, necessitating a split. The principal direction of the density flow field is the direction of the density flow field vector, i.e., (-0.043, 0.023).

[0117] Calculate the difference between the strain energy density and the preset density threshold, 0.085 - 0.08 = 0.005. Substitute this difference into the exponential function f(x) = 1 - exp(-λx) to calculate the splitting coefficient, where λ is typically set to 200. For a difference of 0.005, the splitting coefficient is 1 - exp(-200 × 0.005) = 0.632. The size of the resulting subgrid is determined by the splitting coefficient and the original grid size: the subgrid side length is the original grid side length multiplied by the splitting coefficient. For an original grid with a side length of 10 meters, the subgrid side length after splitting is 10 × 0.632 = 6.32 meters.

[0118] The mesh is split along the main direction of the density flow field, with the splitting line passing through the center of the original mesh and perpendicular to the main direction of the density flow field. For the above mesh cell, the center coordinates are (34.667, 45.179) and the main direction of the density flow field is (-0.043, 0.023). Therefore, the splitting line is perpendicular to this direction, with the direction vector being (0.023, 0.043).

[0119] Calculate the mesh side length ratio, mesh area change rate, and deformation vector difference between adjacent mesh vertices for the split sub-mesh. The mesh side length ratio is the ratio of the longest to shortest side in the sub-mesh and is required to not exceed the preset distortion threshold, typically set to 1.5. The mesh area change rate is the absolute value of the ratio of the sub-mesh area to the original mesh area minus 1. It is required to not exceed the preset area change threshold, typically set to 0.2. The deformation vector difference between adjacent mesh vertices is the modulus of the difference between the deformation vectors of shared vertices in adjacent meshes and is required to not exceed the product of the preset smoothing coefficient and the distance between mesh vertices. The preset smoothing coefficient is typically set to 0.1.

[0120] For example, after the split, the long sides of the two subgrids are 8.5 meters and 8.3 meters, respectively, and the short sides are 6.2 meters and 6.4 meters, respectively. The grid side length ratios are 8.5 / 6.2 = 1.37 and 8.3 / 6.4 = 1.30, respectively, both below the preset distortion threshold of 1.5. The subgrid areas are 52.7 square meters and 53.1 square meters, respectively, and the original grid area is 100 square meters. The grid area change rates are |52.7 / 100-1| = 0.473 and |53.1 / 100-1| = 0.469, respectively, both exceeding the preset area change threshold of 0.2 and failing to meet the requirements.

[0121] At this point, the position or direction of the splitting line needs to be adjusted and the splitting process repeated until all conditions are met. After multiple adjustments, the resulting two sub-grids have areas of 88.7 square meters and 87.2 square meters, respectively, with area change rates of 0.113 and 0.128, respectively, both less than the preset area change threshold of 0.2. The maximum difference in deformation vectors between adjacent grid vertices is 0.85 meters, and the grid vertex distance is 10 meters. The product of the preset smoothing coefficient and the grid vertex distance is 0.1 × 10 = 1 meter, which is less than 1, thus meeting the requirements.

[0122] Therefore, these two subgrids are identified as target grid cells, and their corresponding original point cloud data are recorded to provide a basis for subsequent terrain feature analysis.

[0123] Existing point cloud processing technologies typically use fixed structures like uniform meshing or octrees, which are difficult to adapt to complex terrain variations. For example, traditional uniform meshing methods are ineffective in areas with large variations in point cloud density. They can cause redundant calculations in sparse areas and lose detailed information in dense areas. While methods like octrees have some adaptability, their segmentation scheme is fixed and cannot be optimized based on terrain characteristics.

[0124] The method in this embodiment is based on a density flow field-driven mesh generation method. Starting from the point cloud density distribution, it constructs the core concept of density flow field, enabling meshing to be optimized along the main directions of terrain features. The starting point of the improvement is to improve the adaptability and expressiveness of meshing to terrain features. The density flow field is constructed through a linear combination of density gradient vectors and orthogonal rotation vectors. Strain energy density is introduced to determine the splitting timing, and multiple conditional constraints are used to ensure the effectiveness of the splitting.

[0125] Compared with the existing technology, the main advantages of the method of this embodiment are: introducing the concept of density flow field, so that the grid deformation and splitting can be carried out along the main direction of terrain change; using strain energy density as the splitting criterion, realizing adaptive splitting based on physical meaning; through multiple constraints such as grid side length ratio, area change rate and deformation vector difference, the quality and smooth transition of the split grid are guaranteed.

[0126] The method in this embodiment can more accurately capture terrain features, especially in areas with drastic changes in point cloud density, such as ridges and gullies, where the terrain transitions. The meshing can be adaptively refined, thereby improving the accuracy of subsequent terrain analysis. Experimental results show that compared to traditional uniform grids, this method improves the accuracy of terrain feature preservation by approximately 35% and saves approximately 25% in computational efficiency, providing more reliable data support for path planning and operational decision-making for unmanned systems in complex terrain environments.

[0127] like Figure 2As shown, the simulation results of the relationship between strain energy density and mesh deformation and splitting performance are demonstrated. The horizontal axis represents the strain energy density value, and the vertical axis represents the mesh quality score after deformation (%). The figure shows the performance of the four methods under different strain energy densities. When the strain energy density is 0.085, the mesh quality score of this technical solution reaches 92.7%, which is significantly higher than the 78.4% of the splitting threshold method, 68.9% of the elastic deformation method, and 61.2% of the fixed splitting method. When the strain energy density exceeds the preset density threshold (0.08), this technical solution introduces a density flow field-guided mesh splitting strategy, so that the mesh quality continues to remain at a high level. Even in the extreme case where the strain energy density reaches 0.125, it can still maintain a quality score of 84.8%, while other methods quickly drop to 65.2%, 54.3% and 48.7%. This fully demonstrates the superiority of this technical solution in dealing with high-density gradient areas. In particular, in the strain energy density range of 0.095 to 0.115, the average difference between this solution and the next best method reached 18.6 percentage points. At a strain energy density of 0.105, the mesh quality score of this solution was 89.3%, while the splitting threshold method achieved 72.6%, a gap of 16.7%. The preset density threshold of 0.08, specifically marked in the figure, is a key parameter of this solution. Above this threshold, this solution activates the density flow field main direction splitting algorithm, allowing the mesh quality score to remain high without a significant drop as the strain energy density increases.

[0128] In this embodiment, local refinement is achieved based on point cloud density to enhance the geometric expression ability of the target area; the density gradient is used to construct a density flow field to guide the grid to deform and split preferentially in high strain areas, thereby improving the adaptability of the model to non-uniform data; the grid side length ratio, area change rate and smoothness constraints are introduced to effectively suppress grid distortion and ensure the stability and accuracy of subsequent calculations; the splitting scale is controlled by strain energy density and exponential function, taking into account both local feature capture and global structural rationality.

[0129] In an optional embodiment, calculating the stability index of the crawler vehicle based on the posture data, the track pressure data, and the cleaning arm force data includes:

[0130] Receive attitude data, including pitch angle, roll angle and heading angle; receive track pressure data, including left track pressure distribution data and right track pressure distribution data; receive cleaning arm force data, including torque data and joint axial force data of each joint of the cleaning arm;

[0131] Determining the coordinates of the boundary points of the crawler vehicle support polygon based on the posture data, calculating the distances from the projection point of the crawler vehicle's center of gravity to the sides of the crawler vehicle support polygon, and determining the ratio of the minimum value of the distances to the characteristic size of the crawler vehicle support polygon as the static stability margin;

[0132] Calculating the zero moment point coordinates of the crawler vehicle based on the track pressure data and the cleaning arm force data, and determining the dynamic stability margin as a complementary value of the ratio of the distance between the zero moment point coordinates and the center coordinates of the crawler support polygon to a preset safety radius;

[0133] The weighted sum of the static stability margin and the dynamic stability margin is determined as the tracked vehicle stability index.

[0134] In one specific embodiment, tracked vehicles are often used for cleaning operations in complex terrain environments, and their stability is crucial to operational safety. The method of this embodiment obtains posture data, track pressure data, and cleaning arm force data during the operation of the tracked vehicle. Based on this data, static and dynamic stability margins are calculated. Finally, these two stability margins are weighted and combined to produce a final stability index.

[0135] The crawler is equipped with data acquisition equipment such as attitude sensors, track pressure sensors, and cleaning arm torque sensors. The attitude sensors collect data on the crawler's pitch, roll, and heading angles; the track pressure sensors collect data on the pressure distribution of the left and right tracks; and the cleaning arm torque sensors collect data on the torque and axial force of each joint of the cleaning arm.

[0136] Attitude data collection is achieved through an inertial measurement unit (IMU) mounted on the tracked vehicle. This unit typically integrates an accelerometer, gyroscope, and magnetometer, providing real-time information on the vehicle's attitude relative to the ground. For example, in one actual test, the attitude data obtained included a pitch angle of 5.2 degrees, a roll angle of 3.7 degrees, and a heading angle of 178.5 degrees.

[0137] Track pressure data is collected using a pressure sensor array installed at the track-ground contact area, providing information on the pressure distribution along the length of each track. In practice, eight pressure measurement points are set up on each track, located in key areas at the front, middle, and rear of the track. Example measurement results: The left track pressure distribution is [120, 135, 142, 150, 148, 140, 128, 115] kPa, while the right track pressure distribution is [118, 132, 145, 153, 150, 142, 130, 112] kPa.

[0138] The force data on the cleaning arm is obtained using force / torque sensors installed at each joint of the cleaning arm. For a tracked vehicle equipped with a three-degree-of-freedom cleaning arm, the joint torque data example is [250, 180, 120] N·m, and the joint axial force data example is [800, 650, 400] N.

[0139] Based on the posture data, the coordinates of the boundary points of the tracked vehicle's support polygon are determined. The support polygon is the polygon formed by the contact area between the tracked vehicle and the ground. On ideal flat ground, the support polygon is approximately rectangular. However, in complex terrain, the influence of the tracked vehicle's posture on the support polygon needs to be considered.

[0140] Specifically, a coordinate system for the crawler vehicle is established, with the origin at the vehicle's geometric center, the x-axis pointing toward the vehicle's front, the y-axis pointing to the left, and the z-axis pointing vertically upward. Then, based on the track dimensions (length L = 1.8 meters, width B = 0.6 meters) and the vehicle's posture data, a coordinate transformation is performed to determine the coordinates of the four corner points of the crawler vehicle in the world coordinate system.

[0141] Taking the above pose data as an example, after coordinate transformation, the coordinates of the four boundary points of the support polygon are: the left front corner (-0.9, 0.3, -0.082) meters, the right front corner (-0.9, -0.3, -0.065) meters, the left rear corner (0.9, 0.3, 0.094) meters, and the right rear corner (0.9, -0.3, 0.077) meters. Note that the z coordinate here reflects the height difference caused by the pitch and roll angles.

[0142] Calculate the distance from the projection of the crawler's center of gravity to each edge of the crawler's support polygon. The crawler's center of gravity can be determined using design parameters or obtained through dynamic analysis. In this example, the center of gravity is (0.05, 0, 0.35) meters, and its projection on the horizontal plane is (0.05, 0, 0) meters.

[0143] Calculate the perpendicular distances from the projection point to each side of the supporting polygon, obtaining four distance values: 0.85 meters for the front, 0.3 meters for the right, 0.85 meters for the back, and 0.3 meters for the left. The minimum value, 0.3 meters, is used as the key parameter for determining stability. The ratio of this minimum distance to the characteristic dimension of the supporting polygon (which can be taken as the radius of the inscribed circle of the supporting polygon, 0.3 meters) is used to determine the static stability margin, which in this case is 1.0.

[0144] The coordinates of the crawler vehicle's zero moment point (ZMP) are calculated based on the track pressure data and the cleaning arm force data. The ZMP is the point where the sum of the torques generated by all forces acting on the system is zero, and is an important indicator for evaluating dynamic stability.

[0145] The calculation process first requires integrating the track pressure distribution data to determine the pressure center positions of the left and right tracks. Based on the previous pressure distribution data, the pressure center position of the left track is calculated to be 0.98 meters from the front of the track, and the pressure center position of the right track is 1.02 meters from the front of the track.

[0146] Combined with the force data of the cleaning arm, the coordinates of the zero-moment point were calculated using the principle of moment balance. Considering the contribution of the cleaning arm's posture and force conditions to the overall torque of the system, the calculation results show that the coordinates of the zero-moment point are (0.12, -0.05, 0) meters.

[0147] The distance between the zero moment point coordinates and the support polygon center coordinates (0,0,0) is calculated to be 0.13 meters. The default safety radius is 0.25 meters, and the ratio of the distance between the zero moment point and the polygon center to the safety radius is 0.52. The complementary value of 0.48 is determined as the dynamic stability margin.

[0148] The tracked vehicle stability index is calculated by weighting the static and dynamic stability margins. The weighting coefficients can be adjusted based on actual operational requirements. Typically, a static stability margin weight of 0.6 and a dynamic stability margin weight of 0.4 are used. In this example, the tracked vehicle stability index is calculated as 1.0 × 0.6 + 0.48 × 0.4 = 0.792.

[0149] For the calculated stability index, different thresholds can be set to correspond to different stability states: when the stability index is greater than 0.8, it means that the system is highly stable and can perform high-difficulty operations; when the stability index is between 0.5-0.8, it means that the system is in a normal stable state and is suitable for routine operations; when the stability index is between 0.3-0.5, it means that the system stability is insufficient and the operation intensity needs to be reduced; when the stability index is less than 0.3, it means that the system is in a critical stability state and the operation should be stopped immediately and the posture adjusted.

[0150] This implementation can be further expanded to include consideration of the impact of terrain on supporting polygons, the introduction of filtering algorithms to improve data reliability, and time-series analysis of stability indices to predict potential instability. These expansions can further enhance the accuracy and reliability of tracked vehicle stability assessments and adapt to more complex operating environments.

[0151] In this embodiment, by determining the coordinates of the boundary points of the supporting polygon and calculating the ratio of the minimum distance of the center of gravity projected to each side to the characteristic size of the polygon, the stability margin of the vehicle in the static state is quantified, which provides a basis for evaluating the safety of the vehicle when it is not affected by external dynamics; the zero moment point is calculated in combination with the track pressure and the force data of the cleaning arm, and compared with the coordinates of the center of the supporting polygon, and the dynamic stability margin is determined by using the complement of the distance ratio, so that when the vehicle is subject to external disturbances or load changes, the real-time stability performance can still be evaluated in time; by weighted calculation of the static stability margin and the dynamic stability margin, a comprehensive tracked vehicle stability index is formed, thereby providing an intuitive and quantitative evaluation indicator for the overall stability of the vehicle, which is convenient for monitoring and early warning.

[0152] In an optional embodiment, executing the adaptive entropy enhanced collaborative tracking algorithm to solve the track torque distribution parameters and the cleaning arm posture parameters according to the terrain type, terrain characteristic parameters and the tracked vehicle stability index includes:

[0153] Based on a preset track torque distribution parameter range and a preset cleaning arm posture parameter range, a parameter search space is determined, the parameter search space is divided into multiple equally probable subspaces, and a collaborative tracking agent is generated in each equally probable subspace based on uniformly distributed random sampling. All the collaborative tracking agents generated in the equally probable subspaces constitute a collaborative tracking agent group;

[0154] Calculating a local entropy value for each collaborative tracking agent in the collaborative tracking agent group based on the neighborhood probability density, and determining the mean of the local entropy values as the overall entropy value of the group;

[0155] According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, a fitness function is constructed by weighted combination of terrain evaluation value and stability evaluation value;

[0156] Calculating an adaptive step-size adjustment coefficient based on the overall entropy value of the group, selecting a first agent according to the fitness function, and updating the parameters of the agents in the collaborative tracking agent group in combination with the adaptive step-size adjustment coefficient;

[0157] Calculating the parameter dispersion of the collaborative tracking agent group, determining a group diversity index, and performing Gaussian random mutation when the group diversity index is less than a preset diversity threshold;

[0158] When the change in the optimal parameters of continuous iterations is less than the preset convergence threshold, the current optimal crawler torque distribution parameters and cleaning arm posture parameters are output;

[0159] The weight coefficient of the fitness function is dynamically adjusted based on the terrain type and terrain characteristic parameters updated in real time.

[0160] When a tracked vehicle performs cleaning operations on complex terrain, it is necessary to reasonably distribute the track torque and adjust the cleaning arm posture according to the terrain type, terrain characteristic parameters and stability index to ensure the efficiency and safety of the cleaning operation.

[0161] The parameter search space is determined based on the preset ranges of track torque distribution parameters and cleaning arm posture parameters. The track torque distribution parameter includes the torque ratio TL / TR of the left and right tracks, which is set in the range of [0.5, 2.0]. The cleaning arm posture parameters include the arm's extended length L (range [0.8m, 1.5m]), pitch angle α (range [-30°, 45°]), and yaw angle β (range [-60°, 60°]).

[0162] The parameter search space is divided into multiple equally probable subspaces. Specifically, the torque ratio TL / TR interval [0.5, 2.0] can be evenly divided into 15 subintervals, the extension length L interval [0.8m, 1.5m] can be evenly divided into 10 subintervals, the pitch angle α interval [-30°, 45°] can be evenly divided into 15 subintervals, and the yaw angle β interval [-60°, 60°] can be evenly divided into 20 subintervals, forming a grid division of the multidimensional parameter search space.

[0163] Cooperative tracking agents are randomly sampled from a uniform distribution within each equally probable subspace. For example, within the subspace [1.1, 1.2] × [1.0m, 1.1m] × [0°, 5°] × [-10°, -5°], a cooperative tracking agent with a torque ratio of 1.15, a reach of 1.05m, a pitch angle of 2.5°, and a yaw angle of -7.8° is randomly generated. A total of 200 cooperative tracking agents are generated across all subspaces to form the cooperative tracking agent swarm.

[0164] The local entropy value is calculated for each agent in the collaborative tracking agent group to reflect the uncertainty of the parameter distribution. The local entropy value is calculated as follows: first determine the neighborhood radius r of each agent (for example, take 10% of the length of the diagonal of the parameter space), count the number of agents n within the radius r, and then calculate the probability density p of the area, and then calculate the local entropy value h. For example, there are 8 agents in the neighborhood of an agent, and the total number of agents is 200. Then, the probability density of the area is p = 8 / 200 = 0.04, and the local entropy value h = -p × ln(p) = 0.128. The average local entropy value of all agents is used as the overall entropy value H of the group. In the initial state, the overall entropy value of the group is approximately 0.85.

[0165] A fitness function is constructed based on terrain type, terrain characteristic parameters, and the tracked vehicle stability index. The terrain evaluation value (ET) varies depending on the terrain type: 0.9 for flat terrain, 0.7 for gently sloping terrain, 0.5 for steep slopes, and 0.3 for stepped terrain. Terrain characteristic parameters include the slope angle θ (ranging from 0° to 30°) and the terrain roughness R (ranging from 0 to 1). The stability evaluation value (ES) is calculated based on the tracked vehicle stability index (SI). Higher SI values correspond to higher ES values. The fitness function F is constructed by weightedly combining the terrain evaluation value and the stability evaluation value: F = w1 × ET + w2 × ES, where w1 and w2 are weight coefficients. Initially, w1 = 0.4 and w2 = 0.6.

[0166] The adaptive step-size adjustment coefficient λ is calculated based on the overall entropy value H of the population. When H is large (e.g., H > 0.7), λ is smaller (e.g., 0.3); when H is small (e.g., H < 0.3), λ is larger (e.g., 0.8) to balance global search and local convergence. The agent with the highest fitness value is selected as the first agent, and the remaining agents update their parameters based on the first agent: new parameter value = original parameter value + λ × (first agent parameter value - original parameter value) + random perturbation. For example, if the first agent's torque ratio is 1.3, the torque ratio of another agent is 1.1, λ = 0.5, and the random perturbation is 0.02, then the torque ratio of this agent after the update is 1.1 + 0.5 × (1.3 - 1.1) + 0.02 = 1.22.

[0167] The parameter dispersion of the collaborative tracking agent group is calculated to determine the group diversity index D. The parameter dispersion is obtained by calculating the standard deviation of each parameter in the agent group. The standard deviations of the four parameters are normalized and then weighted averaged to obtain the group diversity index D. When D is less than the preset diversity threshold DT (for example, DT = 0.15), Gaussian random mutation is performed: 30% of the agents are randomly selected and their parameters are perturbed with a Gaussian random perturbation of mean 0 and standard deviation 0.1 to increase the group diversity and prevent it from falling into a local optimum.

[0168] Iterate the above steps. When the change in the optimal parameters for 10 consecutive iterations is less than the preset convergence threshold (for example, the torque ratio change is less than 0.01, the extension length change is less than 0.01m, and the angle change is less than 0.5°), output the current optimal track torque distribution parameters and cleaning arm posture parameters.

[0169] The fitness function's weight coefficients are dynamically adjusted based on real-time updates of terrain type and characteristic parameters. When the terrain is detected as flat or gently sloping, the terrain evaluation value weight w1 is increased (e.g., to 0.6) and the stability evaluation value weight w2 is decreased (e.g., to 0.4). When the terrain is detected as steep or stepped, the terrain evaluation value weight w1 is decreased (e.g., to 0.3) and the stability evaluation value weight w2 is increased (e.g., to 0.7). The weights are also adjusted based on the terrain roughness R: a larger R value results in a larger w2 value.

[0170] For example, a tracked vehicle is operating on a steep slope with a slope angle of 20° and a terrain roughness of 0.6, with an initial stability index of 0.72. After 50 iterations of the adaptive entropy-enhanced collaborative tracking algorithm, the optimal track torque distribution parameters are obtained: torque ratio TL / TR = 1.35 (the left track torque is greater than the right track, which is beneficial to lateral stability); optimal cleaning arm posture parameters: extension length L = 0.95m (a shorter extension length lowers the center of gravity), pitch angle α = 15° (lifted up to adapt to the slope), yaw angle β = -10° (slightly deflected inward to balance the weight). After adopting this set of parameters, the tracked vehicle stability index is improved to 0.86, and the cleaning efficiency is improved by about 25%.

[0171] In this embodiment, the entire parameter search space is divided into multiple equally probable subspaces using the preset track torque distribution parameter range and the cleaning arm posture parameter range, and then the collaborative tracking agents are uniformly sampled in each subspace to effectively cover the global search space and improve the search efficiency and comprehensiveness; the local entropy value under the neighborhood probability density is calculated for each collaborative tracking agent, and its mean is used as the overall entropy value of the group, which can fully reflect the discrete degree of parameter distribution and search diversity within the agent group, thereby providing a basis for subsequent adaptive adjustment of step size and update of parameters; the terrain type, terrain feature parameters and tracked vehicle stability index are weighted and combined to construct a fitness function, so that the search process not only considers the stability of the target equipment, but also takes into account the environmental adaptability, further improving the scientific nature of parameter selection and the reliability of actual application; the adaptive step size adjustment coefficient is calculated according to the overall entropy value of the group, and the agent parameters are dynamically updated. At the same time, the diversity index is used to monitor the discrete degree of group parameters. When the diversity is insufficient, the search space is expanded through Gaussian random mutation to ensure that the algorithm has sufficient exploration ability in the global range and prevent it from falling into the local optimum.

[0172] like Figure 3As shown, the terrain adaptability comparison results of the solution of this embodiment and three existing technologies under different terrain types are demonstrated. As can be seen from the data in the figure, the solution of this embodiment performs well in all tested terrains, especially under complex terrain conditions. In steep slope terrain (slope 42°), the terrain adaptability score of this technical solution reached 87.6 points, which is 12.3 points higher than the parameter adjustment method based on PID control; in soft mud, the adaptability score is 83.2 points, which is 15.7 points higher than the parameter optimization method based on genetic algorithm; on rocky and uneven roads, the adaptability score reaches 92.1 points, which is 18.9 points higher than the adaptive technology based on fuzzy control; in complex terrain, the adaptability score is 85.9 points, which is better than all compared technologies. The data show that the solution of this embodiment uses the adaptive step size adjustment and diversity maintenance mechanism based on the group entropy value to coordinate the optimization of the cleaning arm posture parameters and the track torque distribution, which greatly improves the stability and operating efficiency of the system in various terrains. The figure also shows the iterative convergence speed of different algorithms. The solution of this embodiment reaches a stable state after an average of 15.2 iterations, while the parameter adjustment method based on PID control requires 25.3 iterations, the parameter optimization method based on genetic algorithm requires 27.8 iterations, and the adaptive technology based on fuzzy control requires 28.5 iterations, demonstrating the high efficiency of this solution. Notably, in the area with sudden changes in terrain (marked as area T4 in the figure), the real-time adjustment capability of the solution of this embodiment reduces its adaptability by only 8.3%, which is much lower than the decreases of the parameter adjustment method based on PID control (19.7%), the parameter optimization method based on genetic algorithm (23.4%), and the adaptive technology based on fuzzy control (15.7%).

[0173] In an optional embodiment, a dynamic spatiotemporal constraint cost function is constructed based on the track torque distribution parameter and the terrain characteristic parameter, and an optimal path sequence is obtained through progressive hierarchical prediction time domain planning, including:

[0174] Based on the track torque distribution parameters, the torque distribution cost term is calculated through the left track torque parameters and the right track torque parameters. Based on the terrain characteristic parameters, the terrain characteristic cost term is calculated through the terrain slope parameter and the terrain roughness parameter. The dynamic spatiotemporal constraint cost function is constructed by combining the time cost term and the path smoothness cost term.

[0175] Within the overall planning horizon, a hierarchical prediction horizon architecture is constructed, the prediction horizon length is adaptively adjusted, and the local optimal control sequence that minimizes the dynamic spatiotemporal constraint cost function is solved. The corresponding control instructions are executed and the prediction horizon is updated in a sliding manner. The loop is iterated until the overall planning horizon ends and the planning result is obtained.

[0176] Based on the planning results, an optimal path sequence consisting of position parameters, heading angle parameters, linear velocity parameters, and angular velocity parameters is generated.

[0177] In a specific embodiment, a dynamic spatiotemporal constraint cost function is constructed, which consists of four main parts: a track torque distribution cost term, a terrain feature cost term, a time cost term, and a path smoothness cost term.

[0178] The track torque distribution cost term is calculated using the left and right track torque parameters. Specifically, as the robot moves, the real-time values of the left track torque TL and the right track torque TR are recorded. The torque difference ΔT = |TL-TR| is calculated, and the torque balance coefficient KT (typically 0.7) is set. If the torque difference is greater than the preset threshold Tthreshold (e.g., 50 Nm), the torque distribution cost CT = KT × (ΔT-Tthreshold); otherwise, CT = 0. This design allows the robot to maintain a balanced force on the left and right tracks, improving travel stability.

[0179] The terrain feature cost item comprehensively considers the terrain slope and roughness. The vehicle's onboard sensors obtain the slope angle α (unit: degree) and the terrain roughness index R (range: 0-10, with larger values indicating more uneven terrain). The slope coefficient Kα (typical value: 0.5) and the roughness coefficient KR (typical value: 0.3) are set. The terrain feature cost CF is calculated as CF = Kα × α2 + KR × R. In actual applications, when the slope α exceeds 20 degrees or the roughness R exceeds 8, the terrain feature cost increases significantly, guiding the robot to avoid these areas.

[0180] The time cost term is directly related to the time required for path planning. Setting the time weight coefficient Kt (typically 0.4), the time cost CC = Kt × C, where C is the estimated time to complete the path (in seconds).

[0181] The path smoothing cost term is used to avoid sharp turns along the path. Assuming the heading angles of adjacent path points are θi and θi+1, calculate the angle change Δθ = |θi+1-θi|. Set the smoothing coefficient Ks (typically 0.8), and the smoothing cost CS = Ks × Σ(Δθ)², summed over all pairs of adjacent points on the planned path.

[0182] Combining the above four items, the dynamic spatiotemporal constraint cost function C = CT + CF + CC + CS, the robot will look for the path that minimizes the value of this function.

[0183] A progressive hierarchical prediction time domain planning method is implemented, and a hierarchical architecture is used for path planning within the overall planning time domain.

[0184] Set the total planning time domain Ttotal (for example, 300 seconds) and the initial prediction time domain length Tp_init (for example, 30 seconds). In actual operation, the prediction time domain length will be adaptively adjusted according to the complexity of the environment.

[0185] During each planning cycle, perform the following steps:

[0186] Get the robot's current state, including position coordinates (x, y), heading angle θ, linear velocity v, and angular velocity ω. Also get the surrounding environment information provided by the perception system, including terrain slope, roughness, etc.

[0187] The prediction horizon length, Tp, is adaptively adjusted based on environmental complexity. When increasing environmental complexity is detected (for example, if the terrain feature cost exceeds a preset threshold), the prediction horizon length is shortened to improve local planning accuracy. When the environment is relatively simple, the prediction horizon length is extended to increase planning efficiency. For example, Tp can be set to 50 seconds on flat terrain, while it can be shortened to 20 seconds in areas with a slope greater than 15 degrees.

[0188] Within the current prediction horizon, a model predictive control approach is used to generate a series of feasible control sequences. Each control sequence consists of a set of linear and angular velocity commands {(v1,ω1),(v2,ω2),...,(vn,ωn)}, where n is the number of control steps within the prediction horizon. For each control sequence, the future trajectory is predicted based on the robot's kinematic model, and the corresponding dynamic spatiotemporal constraint cost function is calculated.

[0189] The control sequence that minimizes the cost function value is selected as the local optimal control sequence. In practical applications, the feasibility of the selected path can be ensured by setting a cost function threshold (for example, the total cost is less than 100).

[0190] The robot executes the first control instruction (v1, ω1) in the optimal control sequence for a control period Tc (usually 0.1 seconds).

[0191] Slide and update the prediction time domain, move the prediction window forward by Tc time, and repeat the above steps until the total planning time domain ends.

[0192] Generate an optimal path sequence based on the planning results. The optimal path sequence consists of a series of path points, each of which contains the following parameters:

[0193] Position parameters, including x- and y-coordinates in the global coordinate system, in meters and accurate to centimeters;

[0194] The heading angle parameter indicates the robot's heading direction, ranging from 0 to 360 degrees, with an accuracy of 0.1 degrees;

[0195] Linear velocity parameter, which indicates the robot's forward speed, usually ranges from 0 to 2 m / s and is adaptively adjusted according to terrain conditions;

[0196] Angular velocity parameter, indicating the rotation speed of the robot, usually ranges from -0.5 to 0.5 radians per second. A positive value indicates a right turn, and a negative value indicates a left turn.

[0197] For example, a tracked robot navigates in a mountain environment. The starting coordinates are (0,0) and the end coordinates are (100,50). In the middle, it needs to cross a slope area with a slope of 18 degrees and a gravel road with a roughness index of 7. The optimal path planned by this method avoids the steepest slope, chooses an area with a gentler slope (about 12 degrees) to climb, and reduces the travel speed in the gravel road area (from 1.2 m / s to 0.8 m / s). The total length of the planned path is 158 meters, and the estimated travel time is 210 seconds. The difference in torque between the left and right tracks is kept below 40 Nm throughout the entire process, ensuring travel stability.

[0198] Experimental results show that compared to the traditional A* algorithm, this method reduces path execution time by 15% and energy consumption by 22%, and significantly improves travel stability in areas with varying slopes. This method is particularly suitable for path planning in complex environments that require comprehensive consideration of terrain characteristics and robot dynamics.

[0199] In an optional embodiment, within the overall planning horizon, a hierarchical prediction horizon architecture is constructed, the prediction horizon length is adaptively adjusted, a local optimal control sequence is solved that minimizes the dynamic spatiotemporal constraint cost function, corresponding control instructions are executed, and the prediction horizon is updated in a sliding manner. The loop is iterated until the overall planning horizon ends, and the planning results obtained include:

[0200] The overall planning time domain is divided into a macro-planning layer time domain and a micro-execution layer time domain, wherein the macro-planning layer time domain includes multiple consecutive prediction cycles, and each prediction cycle corresponds to a micro-execution layer time domain;

[0201] Collecting environmental information of the current location to obtain environmental complexity, detecting a deviation between the actual trajectory and the planned trajectory, determining a length of a predicted time domain based on the environmental complexity, and determining an adjustment interval for the predicted time domain based on the deviation;

[0202] Dividing the current prediction time domain into two adjacent time intervals according to a preset ratio, determining a first time interval and a second time interval, establishing a dynamic prediction model including track-ground interaction forces to generate a motion trajectory in the first time interval, establishing a kinematic prediction model based on steering constraints to generate a path direction in the second time interval, and combining the motion trajectory and the path direction to generate a predicted trajectory;

[0203] Based on the predicted trajectory, a control sequence is solved to minimize the dynamic spatiotemporal constraint cost function, the control instruction corresponding to the first time interval is executed, the predicted time domain is updated to the starting point of the next prediction, and the iterative solution is continued until the total planning time domain ends to obtain the planning result.

[0204] In one specific embodiment, a time-domain partitioning structure is designed to divide the total time domain required for vehicle planning into two levels: a macro-planning layer time domain and a micro-execution layer time domain. The macro-planning layer time domain is typically the execution time of the entire task, for example, a 300-second drive from the starting point to the end point. This time domain is divided into multiple consecutive prediction cycles, and the length of each prediction cycle is dynamically adjusted based on the complexity of the environment. For example, in a simple flat environment, the prediction cycle can be set to 10 seconds, while in a complex and rugged environment, it can be shortened to 5 seconds. Each prediction cycle corresponds to a micro-execution layer time domain, which is responsible for generating and executing specific control instructions.

[0205] The vehicle uses onboard LiDAR, visual sensors, and other devices to collect environmental information about the current location and calculate an environmental complexity index. Environmental complexity can be comprehensively assessed based on factors such as terrain undulation and obstacle density, and is represented by a normalized value between 0 and 1. For example, the environmental complexity of a flat, open area is 0.1, while the environmental complexity of a steep area with many obstacles can reach 0.8. At the same time, the deviation between the actual driving trajectory and the planned trajectory of the previous cycle is detected, and the weighted sum of the lateral position deviation and the heading angle deviation is calculated. For example, if the lateral position deviation at a certain moment is 0.4 meters and the heading angle deviation is 5 degrees (approximately 0.087 radians), and the weights are 1 and 5, respectively, then the weighted deviation value is 0.4 + 5 × 0.087 = 0.835.

[0206] The length of the prediction time domain is determined based on the complexity of the environment. The length of the prediction time domain is inversely proportional to the complexity of the environment. It can be obtained by multiplying the preset basic prediction time domain by the inverse function of the environmental complexity. Assuming that the basic prediction time domain is 10 seconds and the environmental complexity is 0.8, the actual prediction time domain length is 10×(1-0.8+0.2)=4 seconds. The adjustment interval of the prediction time domain is determined based on the deviation value. The larger the deviation value, the more frequent the adjustment. For example, when the deviation value is less than 0.5, the prediction time domain is adjusted every 2 seconds; when the deviation value is between 0.5 and 1, it is adjusted every 1 second; when the deviation value is greater than 1, it is adjusted every 0.5 seconds.

[0207] The current prediction horizon is divided into two adjacent time intervals according to a preset ratio. The preset ratio is usually set to 3:7 or 4:6, that is, the first 30%-40% of the time is the first time interval, and the last 60%-70% of the time is the second time interval. For example, if the prediction horizon is 4 seconds, the first time interval is the first 1.2 seconds (30%), and the second time interval is the last 2.8 seconds (70%).

[0208] During the first time interval, a dynamic prediction model is established that includes track-ground forces. This model considers parameters such as vehicle mass, moment of inertia, friction coefficient between the track and the ground, and terrain slope. It calculates physical quantities such as left and right track drive torque, traction, and lateral force. Based on Newton's second law, it predicts the vehicle's acceleration, velocity, and position changes, generating a highly accurate motion trajectory. For example, on a dirt road with a slope of 15 degrees and a ground friction coefficient of 0.4, a tracked vehicle with a mass of 500 kg and drive torques of 300 Nm and 280 Nm, respectively, will move forward 2.3 meters and deviate 0.15 meters to the left within 1.2 seconds, as calculated by the dynamic model.

[0209] During the second time interval, a kinematic prediction model based on steering constraints is established. This model simplifies the vehicle's dynamic characteristics, primarily considering kinematic characteristics such as the vehicle's turning radius constraint and maximum speed constraint. It predicts the vehicle's future position and orientation based on its current speed and angular velocity, generating a path over a longer timescale. For example, if the vehicle's current speed is 2.5 m / s, its angular velocity is 0.1 rad / s, and its minimum turning radius is 2 meters, then over the next 2.8 seconds, the vehicle will move along an arc with a radius of approximately 25 meters, with a position change of 7 meters forward and 0.49 meters to the right.

[0210] The motion trajectory generated in the first time interval and the path direction generated in the second time interval are combined to form a complete predicted trajectory. When combining, the position, velocity, and heading angle of the two parts at the intersection are guaranteed to be continuous, and smoothing is performed if necessary. For example, for the above calculation results, at 1.2 seconds, the vehicle position is 2.3 meters forward, 0.15 meters to the left, the speed is 2.5 meters / second, and the heading angle is slightly deflected to the right. This matches the initial conditions of the model in the second time interval. After combination, the complete predicted trajectory is that the vehicle will move forward 9.3 meters in 4 seconds and eventually deviate to the right by about 0.34 meters.

[0211] Based on the predicted trajectory, a control sequence is solved. A dynamic spatiotemporal constraint cost function is established, comprising a weighted sum of multiple evaluation metrics such as trajectory tracking error, control energy consumption, and ride smoothness. Model predictive control methods are used to find the control sequence that minimizes this cost function. For example, for a tracked vehicle, the control sequence includes drive torque or speed commands for the left and right tracks. In one solution, the optimal control sequence was a left track torque sequence of [320, 310, 305, 300] Nm and a right track torque sequence of [290, 295, 300, 300] Nm, corresponding to four control time points.

[0212] Execute the control instructions corresponding to the first time interval. For example, for the first 1.2-second time interval, execute the first 30% of the control sequence, which specifies 320 Nm of torque for the left track and 290 Nm for the right track. After execution, update the prediction horizon to the starting point of the next prediction, moving it forward 1.2 seconds. Recollect environmental information, adjust the prediction horizon length, build the prediction model, solve the control sequence, and execute the control instructions. This iterative cycle continues until the navigation task within the total planned time interval is completed.

[0213] Existing tracked vehicle navigation and control methods fall into two main categories: path tracking methods based on simplified kinematic models, which require minimal computation but suffer from low accuracy; and trajectory planning methods based on complete dynamic models, which offer high accuracy but high computational complexity, making them difficult to implement in real time. The former ignores important physical properties such as track-ground interaction, resulting in insufficient control accuracy in complex terrain. While the latter incorporates detailed physical models, it is computationally intensive, making long-term planning difficult and poorly adaptable to environmental changes.

[0214] The method of this embodiment starts from the perspective of time domain division, proposes a two-layer time domain structure of macro planning layer and micro execution layer, and adopts a dynamic and kinematic hybrid prediction model in the prediction domain. The starting point of the improvement is to balance the contradiction between control accuracy and computational efficiency. By using a high-precision dynamic model in a short time interval and an efficient kinematic model in a long time interval, the requirements of accuracy and efficiency are taken into account. At the same time, through the dual feedback mechanism of environmental complexity and trajectory deviation, adaptive adjustment of the prediction domain is achieved, which improves the robustness of the method in complex and changing environments.

[0215] Compared with the existing technology, the advantages of the method of this embodiment are mainly reflected in: first, it introduces an adaptive time-domain adjustment mechanism based on environmental complexity and trajectory deviation, so that the system can flexibly adjust the planning strategy according to actual conditions; second, it adopts a hybrid prediction model of dynamics and kinematics to ensure high-precision control in the critical short time domain and ensure computational efficiency in the long time domain; third, through the model predictive control method, multiple constraints and optimization objectives are integrated into a unified spatiotemporal constraint cost function, realizing the generation of a globally optimal control sequence.

[0216] The method of this embodiment significantly improves computational efficiency and environmental adaptability while maintaining control accuracy. Experimental results show that compared with traditional single-model methods, this method improves trajectory tracking accuracy by approximately 40% and reduces computational time by approximately 60% in complex terrain environments. The improvement in navigation performance is particularly significant in areas with sudden changes in terrain conditions, providing effective technical support for autonomous navigation of tracked vehicles in complex environments.

[0217] like Figure 4The figure shows the adaptive time-domain adjustment effect of this embodiment during a 300-meter navigation process for a tracked vehicle. The figure clearly divides the navigation path into five characteristic areas: a flat area (0-40m), an uphill and rugged area (40-100m), a gentle slope transition area (100-180m), a multi-obstacle complex area (180-270m), and a gentle end area (270-300m). The data shows how terrain complexity and trajectory deviation affect the dynamic changes in the prediction horizon length and adjustment interval: in the flat area (0-40m), the terrain complexity remains at a low level (0.15-0.22) and the trajectory deviation is small (0.12-0.25m). The system adopts a longer prediction horizon (8.24-8.64s) and a larger adjustment interval (2.0s). After entering the uphill and rugged area, the terrain complexity rises sharply to 0.89 and the trajectory deviation increases to 0.92m. The system adaptively shortens the prediction horizon to 3.12s and reduces the adjustment interval to 0.5s, reflecting the system's agile response to environmental changes. In the multi-obstacle complex area (180-270m), the extreme cases of maximum terrain complexity of 0.91 and trajectory deviation of 1.32m occur. At this time, the prediction horizon is shortened to a minimum of 2.92s, and the adjustment interval is maintained at a minimum of 0.5s, ensuring that the system can frequently adjust the control strategy to cope with complex environments.

[0218] The chart distinguishes areas with different terrain characteristics through clear color blocks and displays the average value of key performance indicators in each area, making the relationship between terrain characteristics and control parameters clear at a glance. When the terrain complexity increases from 0.18 in flat areas to 0.82 in multi-obstacle complex areas, the prediction time domain decreases from 8.42s to 3.58s, a decrease of 57.5%; at the same time, when the trajectory deviation increases from 0.18m to 1.06m, the adjustment interval decreases from 2.0s to 0.64s, a decrease of 68%. Two key areas are specially marked in the figure: the area with a sharp increase in slope (complexity peak 0.89) and the area with maximum trajectory deviation (deviation peak 1.32m), and the location of these key points and the corresponding control parameters are intuitively displayed through dotted line indicators.

[0219] The adaptive nature of this technical solution is fully demonstrated in the chart: as environmental complexity increases, the system automatically shortens the prediction horizon to improve control accuracy; as trajectory deviation increases, the system automatically reduces the adjustment interval to increase control frequency; and in transitional regions, the system smoothly adjusts the control strategy to avoid sudden changes. This dual feedback mechanism, based on environmental complexity and trajectory deviation, enables the system to adopt a long-term, low-frequency planning strategy to improve efficiency in simple environments, and a short-term, high-frequency planning strategy to ensure accuracy in complex environments. This fully demonstrates the adaptive nature and environmental perception capabilities of this technical solution, providing strong technical support for the reliable navigation of tracked vehicles in complex and changing environments.

[0220] The embodiment of the present invention is based on a multi-terrain photovoltaic cleaning crawler vehicle stability adaptive control system comprising:

[0221] The first unit is used to obtain the posture data, track pressure data, cleaning arm force data and terrain scanning data of the photovoltaic cleaning crawler vehicle to construct a three-dimensional point cloud model of the terrain;

[0222] The second unit is used to adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain feature parameters based on multi-feature fusion;

[0223] The third unit is used to calculate the stability index of the crawler vehicle based on the posture data, the crawler pressure data and the cleaning arm force data;

[0224] The fourth unit is used to execute the adaptive entropy enhanced collaborative tracking algorithm to solve the track torque distribution parameters and the cleaning arm posture parameters according to the terrain type, terrain characteristic parameters and the tracked vehicle stability index;

[0225] The fifth unit is used to construct a dynamic spatiotemporal constraint cost function based on the track torque distribution parameters and terrain characteristic parameters, and obtain the optimal path sequence through progressive rolling horizon planning;

[0226] The sixth unit is used to generate crawler vehicle travel instructions and cleaning arm posture instructions based on the optimal path sequence and cleaning arm posture parameters in combination with the predictive control strategy to achieve stable operation of the photovoltaic cleaning crawler vehicle.

[0227] According to a third aspect of the embodiments of the present invention,

[0228] An electronic device is provided, comprising:

[0229] processor;

[0230] a memory for storing processor-executable instructions;

[0231] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0232] According to a fourth aspect of the embodiments of the present invention,

[0233] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0234] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0235] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for adaptive control of stability of a multi-terrain photovoltaic cleaning crawler vehicle, characterized in that: include: Obtain the photovoltaic cleaning crawler vehicle's posture data, crawler pressure data, cleaning arm force data, and terrain scanning data to construct a three-dimensional point cloud model of the terrain; Adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain feature parameters based on multi-feature fusion; Calculate the crawler vehicle stability index based on posture data, crawler pressure data and cleaning arm force data; According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, an adaptive entropy enhanced collaborative tracking algorithm is implemented to solve the track torque distribution parameters and cleaning arm posture parameters. A dynamic spatiotemporal constraint cost function is constructed based on track torque distribution parameters and terrain characteristic parameters, and the optimal path sequence is obtained through progressive rolling horizon planning. According to the optimal path sequence and cleaning arm posture parameters, the crawler vehicle travel instructions and cleaning arm posture instructions are generated in combination with the predictive control strategy to achieve stable operation of the photovoltaic cleaning crawler vehicle.

2. The method according to claim 1, characterized in that Adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain characteristic parameters based on multi-feature fusion, including: Project the terrain 3D point cloud model onto the horizontal plane, calculate the initial side length based on the total area and total number of point clouds, and divide it into multiple initial grid units; Calculate the point cloud density distribution of the initial grid cell, construct the density flow field based on the density gradient, and calculate the deformation vector to generate the deformed grid. When the strain energy density of the deformed grid exceeds the preset density threshold, split along the main direction of the density flow field to generate the target grid cell; Calculate the height range within the target grid cell to determine the height difference value, calculate the height change rate in the orthogonal direction to determine the height gradient; establish the height correlation characteristics between target grid cells, calculate the height difference coefficient of adjacent cells, and construct a spatial correlation matrix; The height difference value, height gradient and spatial correlation matrix are combined into a feature vector, which is input into the multi-feature fusion terrain classification discriminant function to determine the terrain type of the target grid cell based on the conditional probability maximization criterion; The terrain slope of the target grid cell is calculated based on the height gradient, and the terrain roughness is determined by calculating the integral deviation between the actual height and the least squares fitting plane; The terrain type, terrain slope and terrain roughness are combined into a terrain characteristic parameter vector to determine the terrain characteristic parameters of the target grid cell.

3. The method according to claim 2, characterized in that Calculate the point cloud density distribution of the initial grid cell, construct the density flow field based on the density gradient, and calculate the deformation vector to generate the deformed grid. When the strain energy density of the deformed grid exceeds the preset density threshold, it is split along the main direction of the density flow field to generate the target grid cells including: The number of point clouds in the initial grid cells is counted, and the number of point clouds per unit area is calculated to obtain the point cloud density distribution. The density change rate in the X direction and the density change rate in the Y direction are calculated according to the point cloud density of adjacent initial grid cells to form a density gradient vector. Perform an orthogonal rotation on the density gradient vector to obtain an orthogonal rotation vector, perform a linear combination of the density gradient vector and the orthogonal rotation vector to construct a density flow field, and calculate the grid deformation vector according to the direction of the density flow field and the modulus of the density gradient vector; Apply the mesh deformation vector to the vertices of the initial mesh unit to generate a deformed mesh unit, and calculate the deformation work per unit area in the deformed mesh unit to obtain the strain energy density; When the strain energy density exceeds a preset density threshold, the deformed grid unit is split along the main direction of the density flow field, the difference between the strain energy density and the preset density threshold is calculated, and the difference is substituted into an exponential function to determine the size of the split subgrid; The mesh side length ratio, mesh area change rate, and deformation vector difference of adjacent mesh vertices of the split sub-mesh are calculated. When the mesh side length ratio does not exceed the preset distortion threshold, the mesh area change rate does not exceed the preset area change threshold, and the deformation vector difference does not exceed the product of the preset smoothing coefficient and the mesh vertex distance, the target mesh unit is determined.

4. The method according to claim 1, wherein The crawler vehicle stability index is calculated based on posture data, track pressure data and cleaning arm force data, including: Receive attitude data, including pitch angle, roll angle and heading angle; receive track pressure data, including left track pressure distribution data and right track pressure distribution data; receive cleaning arm force data, including torque data and joint axial force data of each joint of the cleaning arm; Determining the coordinates of the boundary points of the crawler vehicle support polygon based on the posture data, calculating the distances from the projection point of the crawler vehicle's center of gravity to the sides of the crawler vehicle support polygon, and determining the ratio of the minimum value of the distances to the characteristic size of the crawler vehicle support polygon as the static stability margin; Calculating the zero moment point coordinates of the crawler vehicle based on the track pressure data and the cleaning arm force data, and determining the dynamic stability margin as a complementary value of the ratio of the distance between the zero moment point coordinates and the center coordinates of the crawler support polygon to a preset safety radius; The weighted sum of the static stability margin and the dynamic stability margin is determined as the tracked vehicle stability index.

5. The method according to claim 1, wherein According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, the adaptive entropy enhanced collaborative tracking algorithm is executed to solve the track torque distribution parameters and cleaning arm posture parameters, including: Based on a preset track torque distribution parameter range and a preset cleaning arm posture parameter range, a parameter search space is determined, the parameter search space is divided into multiple equally probable subspaces, and a collaborative tracking agent is generated in each equally probable subspace based on uniformly distributed random sampling. All the collaborative tracking agents generated in the equally probable subspaces constitute a collaborative tracking agent group; Calculating a local entropy value for each collaborative tracking agent in the collaborative tracking agent group based on the neighborhood probability density, and determining the mean of the local entropy values as the overall entropy value of the group; According to the terrain type, terrain characteristic parameters and tracked vehicle stability index, a fitness function is constructed by weighted combination of terrain evaluation value and stability evaluation value; Calculating an adaptive step-size adjustment coefficient based on the overall entropy value of the group, selecting a first agent according to the fitness function, and updating the parameters of the agents in the collaborative tracking agent group in combination with the adaptive step-size adjustment coefficient; Calculating the parameter dispersion of the collaborative tracking agent group, determining a group diversity index, and performing Gaussian random mutation when the group diversity index is less than a preset diversity threshold; When the change in the optimal parameters of continuous iterations is less than the preset convergence threshold, the current optimal crawler torque distribution parameters and cleaning arm posture parameters are output; The weight coefficient of the fitness function is dynamically adjusted based on the terrain type and terrain characteristic parameters updated in real time.

6. The method according to claim 1, characterized in that Based on the track torque distribution parameters and terrain characteristic parameters, a dynamic spatiotemporal constraint cost function is constructed. The optimal path sequence is obtained through progressive hierarchical prediction time domain planning, including: Based on the track torque distribution parameters, the torque distribution cost term is calculated through the left track torque parameters and the right track torque parameters. Based on the terrain characteristic parameters, the terrain characteristic cost term is calculated through the terrain slope parameter and the terrain roughness parameter. The dynamic spatiotemporal constraint cost function is constructed by combining the time cost term and the path smoothness cost term. Within the overall planning horizon, a hierarchical prediction horizon architecture is constructed, the prediction horizon length is adaptively adjusted, and the local optimal control sequence that minimizes the dynamic spatiotemporal constraint cost function is solved. The corresponding control instructions are executed and the prediction horizon is updated in a sliding manner. The loop is iterated until the overall planning horizon ends and the planning result is obtained. Based on the planning results, an optimal path sequence consisting of position parameters, heading angle parameters, linear velocity parameters, and angular velocity parameters is generated.

7. The method according to claim 6, characterized in that Within the overall planning horizon, a hierarchical prediction horizon architecture is constructed, the prediction horizon length is adaptively adjusted, and the local optimal control sequence that minimizes the dynamic spatiotemporal constraint cost function is solved. The corresponding control instructions are executed and the prediction horizon is updated in a sliding manner. The loop is iterated until the overall planning horizon ends. The planning results include: The overall planning time domain is divided into a macro-planning layer time domain and a micro-execution layer time domain, wherein the macro-planning layer time domain includes multiple consecutive prediction cycles, and each prediction cycle corresponds to a micro-execution layer time domain; Collecting environmental information of the current location to obtain environmental complexity, detecting a deviation between the actual trajectory and the planned trajectory, determining a length of a predicted time domain based on the environmental complexity, and determining an adjustment interval for the predicted time domain based on the deviation; Dividing the current prediction time domain into two adjacent time intervals according to a preset ratio, determining a first time interval and a second time interval, establishing a dynamic prediction model including track-ground interaction forces to generate a motion trajectory in the first time interval, establishing a kinematic prediction model based on steering constraints to generate a path direction in the second time interval, and combining the motion trajectory and the path direction to generate a predicted trajectory; Based on the predicted trajectory, a control sequence is solved to minimize the dynamic spatiotemporal constraint cost function, the control instruction corresponding to the first time interval is executed, the predicted time domain is updated to the starting point of the next prediction, and the iterative solution is continued until the total planning time domain ends to obtain the planning result.

8. A multi-terrain photovoltaic cleaning crawler vehicle stability adaptive control system for implementing the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the posture data, track pressure data, cleaning arm force data and terrain scanning data of the photovoltaic cleaning crawler vehicle to construct a three-dimensional point cloud model of the terrain; The second unit is used to adaptively mesh the terrain 3D point cloud model to obtain multiple grid cells, establish highly correlated features between grid cells, and determine the terrain type and terrain feature parameters based on multi-feature fusion; The third unit is used to calculate the stability index of the crawler vehicle based on the posture data, the crawler pressure data and the cleaning arm force data; The fourth unit is used to execute the adaptive entropy enhanced collaborative tracking algorithm to solve the track torque distribution parameters and the cleaning arm posture parameters according to the terrain type, terrain characteristic parameters and the tracked vehicle stability index; The fifth unit is used to construct a dynamic spatiotemporal constraint cost function based on the track torque distribution parameters and terrain characteristic parameters, and obtain the optimal path sequence through progressive rolling horizon planning; The sixth unit is used to generate crawler vehicle travel instructions and cleaning arm posture instructions based on the optimal path sequence and cleaning arm posture parameters in combination with the predictive control strategy to achieve stable operation of the photovoltaic cleaning crawler vehicle.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images