Intelligent control method and system for obstacle avoidance of aerial work platform

Abstract

The invention discloses an intelligent obstacle avoidance control method and system for an aerial work platform, and relates to the field of aerial work platforms, and the method comprises the steps: collecting the sensing data of a working area of the aerial work platform, and constructing a three-dimensional space map based on the sensing data; identifying dynamic and static obstacles in the three-dimensional space map, and performing risk grade division on the working area in combination with a topological potential field algorithm; according to a working area division result, planning an initial obstacle avoidance path by using a random tree algorithm, and optimizing the initial obstacle avoidance path in combination with construction parameters of the aerial work platform to generate a final obstacle avoidance path; and decoupling the final obstacle avoidance path into a control instruction of vertical lifting and horizontal displacement, executing a corresponding obstacle avoidance behavior mode based on the control instruction, and monitoring an execution result to optimize the control instruction. According to the invention, by constructing the three-dimensional space map, behavior modeling of the dynamic obstacle is realized, and the safety and efficiency of path planning are improved.

Description

Technical Field

[0001] This invention relates to the field of aerial work platforms, and more specifically, to an intelligent control method and system for obstacle avoidance of aerial work platforms. Background Technology

[0002] An aerial work platform (AWP), also commonly known as an aerial work vehicle or sling platform, is a type of equipment specifically designed for working at heights. These platforms are widely used in construction, power maintenance, bridge inspection, billboard maintenance, and warehousing operations. AWPs provide workers with a safe and stable platform to perform various tasks at height, such as building facade repair, equipment installation, and cleaning. Due to the unique nature of working at heights, the operating environment is often complex and variable, filled with various obstacles and potentially hazardous areas, making ensuring worker safety a significant challenge.

[0003] Traditional aerial work platforms mostly rely on manual operation, with operators avoiding obstacles and preventing accidents by observing their surroundings. While effective, this method has significant limitations. First, operators' attention is easily affected by fatigue, work hours, complex environments, or unexpected situations, potentially causing them to overlook potential obstacles or risks and leading to accidents. Furthermore, due to the limitations of the human eye in judging distance, angle, and space, the success rate and accuracy of manual obstacle avoidance are low in narrow or crowded environments.

[0004] Currently, obstacle avoidance technology for aerial work platforms has made some progress, mainly relying on sensor detection and basic obstacle avoidance algorithms to achieve automated obstacle avoidance. Many devices are equipped with sensors such as ultrasonic sensors and lidar to detect surrounding obstacles. However, these sensors have certain limitations in complex environments; their detection range and accuracy are limited, and they are easily affected by environmental factors such as lighting, weather changes, or surface reflections. Furthermore, traditional obstacle avoidance algorithms are usually based on preset working environment models, primarily optimized for static obstacles and relatively simple terrain. For dynamic obstacles and complex, constantly changing terrain, these algorithms have poor adaptability and response capabilities, often failing to provide comprehensive and efficient obstacle avoidance protection.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes an intelligent control method and system for obstacle avoidance of aerial work platforms, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, an intelligent control method for obstacle avoidance of an aerial work platform is provided, the method comprising: S1. Collect sensing data of the working area of ​​the aerial work platform and construct a three-dimensional spatial map based on the sensing data; S2. Identify dynamic and static obstacles in the 3D spatial map, and classify the risk level of the work area using the topological potential field algorithm; S3. Based on the work area division results, the initial obstacle avoidance path is planned using the random tree algorithm. The initial obstacle avoidance path is then optimized by combining the construction parameters of the aerial work platform to generate the final obstacle avoidance path. S4. Decouple the final obstacle avoidance path into control commands for vertical lifting and horizontal displacement, execute the corresponding obstacle avoidance behavior mode based on the control commands, and monitor the execution results to optimize the control commands. S2 includes: S21. Based on a three-dimensional spatial map, use deep learning algorithms to identify and classify dynamic and static obstacles within the working area, and obtain the location information of dynamic and static obstacles; S22. Use digital twin technology to construct a digital twin model and combine it with potential field analysis algorithm to predict the motion trajectory of dynamic obstacles; S23. Combining the prediction results of dynamic obstacles with the location information of static obstacles, the working area is divided into a passable area, a cautious area, and a prohibited area using the topological potential field algorithm to identify different dynamic risk levels.

[0008] Optionally, collecting sensing data of the working area of ​​the aerial work platform and constructing a three-dimensional spatial map based on the sensing data includes: S11. Collect sensing data of the working area of ​​the aerial work platform through multimodal sensors, and perform spatiotemporal synchronization and coordinate alignment on the collected sensing data; S12. The Kalman filter algorithm is used to fuse the sensing data after spatiotemporal synchronization and coordinate alignment, and the density-based clustering algorithm is combined to denoise the fused sensing data. S13. Divide the denoised perception data into a voxel grid, extract the centroid of each voxel as the target point, and construct a three-dimensional spatial map by stitching together the target points.

[0009] Optionally, a digital twin model can be constructed using digital twin technology, and the motion trajectory of dynamic obstacles can be predicted by combining it with a potential field analysis algorithm, including: S221. Based on the location information of dynamic obstacles, a digital twin model is constructed using digital twin technology, and the motion state of the dynamic obstacles is updated through the fused real-time perception data. S222. Perform time-frequency analysis on motion perception data in the digital twin model, extract motion cycle features, and combine graph neural networks to mine the interactive topological relationships between dynamic obstacles to generate a spatiotemporal correlation risk matrix to quantify dynamic trends and potential risks. S223. Map the spatiotemporal correlation risk matrix to a potential field intensity distribution, use the fast multipole algorithm to calculate the potential field changes caused by the motion of dynamic obstacles, and establish a kinematic model of dynamic obstacles based on the potential field change results. S224. Combining kinematic models with long short-term memory networks, predict the trajectory of obstacles and optimize the kinematic model through real-time perception data feedback.

[0010] Optionally, time-frequency analysis is performed on the motion perception data in the digital twin model to extract motion cycle features, and graph neural networks are used to mine the interaction topological relationships between dynamic obstacles to generate a spatiotemporal correlation risk matrix to quantify dynamic trends and potential risks, including: S2221. Use short-time Fourier transform to perform time-frequency analysis on motion sensing data in digital twin model, extract periodic features of motion sensing data, and obtain motion periodic features of dynamic obstacles. S2222. Based on the motion perception data of dynamic obstacles, each dynamic obstacle is defined as a node, and the interaction relationship between obstacles is defined as an edge, thus constructing an interaction topology graph; S2223. Use graph neural networks to learn the interaction topology graph, identify and analyze the topological structure and dynamic interaction patterns between dynamic obstacles, and capture the spatiotemporal interaction information between dynamic obstacles. S2224. Combining motion cycle characteristics and spatiotemporal interaction information, calculate the potential risks between dynamic obstacles, construct a spatiotemporal correlation risk matrix, and predict the dynamic changes and potential risks of obstacles by analyzing the changing trend of the spatiotemporal correlation risk matrix.

[0011] Optionally, the spatiotemporal correlation risk matrix is ​​mapped to a potential field intensity distribution, and the potential field changes caused by the motion of dynamic obstacles are calculated using the fast multipole algorithm. Based on the potential field change results, a kinematic model of the dynamic obstacles is established, including: S2231. Use a nonlinear mapping function to convert the spatiotemporal correlation risk matrix into a potential field intensity distribution, and assign the mapped potential field intensity values ​​to the corresponding grid nodes to construct a potential field intensity distribution map. S2232. Use an octree structure to spatially divide the potential field intensity distribution map, and perform cluster analysis based on the location information of dynamic obstacles to form a tree-like hierarchical structure. S2233. Combining the tree-like hierarchical structure with real-time motion perception data of dynamic obstacles, the fast multipole algorithm is used to calculate the potential field changes caused by the motion of dynamic obstacles and update the potential field intensity distribution. S2234. Extract the motion characteristics of the dynamic obstacle from the updated potential field intensity distribution, and establish a kinematic model of the dynamic obstacle based on the motion characteristics.

[0012] Alternatively, the formula for the kinematic model is: ; In the formula, r ( t +1) indicates that the dynamic obstacle occurs in time. t +1 position; r ( t ) indicates the dynamic obstacle in time t Location; t Indicates time; v ( t ) indicates the dynamic obstacle in time t speed; Δ t Indicates the time step; F ( t ) indicates the dynamic obstacle in time t The external force caused by the potential field it experiences; m This indicates the mass of a dynamic obstacle.

[0013] Optionally, based on the work area division results, an initial obstacle avoidance path is planned using a random tree algorithm. This initial path is then optimized by combining it with the construction parameters of the aerial work platform to generate the final obstacle avoidance path, which includes: S31. Based on the region division results, construct a three-dimensional grid map and use a Gaussian mixture model to fit the obstacle distribution boundary to generate obstacle avoidance constraint boundary. S32. Based on the obstacle avoidance constraint boundary, use the target bias-based random tree algorithm to perform path search and generate the initial obstacle avoidance path; S33. Obtain the construction parameters of the aerial work platform, define an optimization target set including path length, minimum safe distance, curvature smoothness and energy consumption coefficient, and use the entropy weight algorithm to dynamically adjust the weight of the optimization target according to the control task requirements. S34. Based on the weight adjustment results, the initial obstacle avoidance path is adjusted using the particle swarm optimization algorithm to generate the final obstacle avoidance path.

[0014] Optionally, the final obstacle avoidance path is decoupled into control commands for vertical lifting and horizontal displacement. Based on these control commands, corresponding obstacle avoidance behavior modes are executed, and the execution results are monitored to optimize the control commands. S41. Based on the preset spatial coordinates, the final obstacle avoidance path is decoupled into a sequence of vertical and horizontal coordinates, and a polynomial fitting technique is used to generate the motion profile of the aerial work platform. S42. Analyze the motion profile, use fuzzy control algorithm to generate control commands for vertical lifting and horizontal displacement, and accelerate the generation of control commands through a parallel computing framework; S43. Select and execute the corresponding obstacle avoidance behavior mode according to the generated control command. The obstacle avoidance behavior mode includes emergency braking mode and dynamic detour behavior mode. S44. Monitor the execution effect of the selected obstacle avoidance behavior mode in real time, and optimize control commands based on the monitoring results.

[0015] Optionally, the motion profile is analyzed, and fuzzy control algorithms are used to generate control commands for vertical lifting and horizontal displacement. The generation of these control commands is accelerated using a parallel computing framework, including: S421. Extract vertical and horizontal feature parameters from the generated motion profile to establish a fuzzy input space. The feature parameters include height deviation, velocity and curvature error. S422. Combining fuzzy input space and fuzzy control algorithm, a fuzzy control model is constructed, which includes an input layer, a membership degree calculation layer and a fuzzy inference layer. A rule reduction layer is introduced to optimize the fuzzy control model. S423. The optimized fuzzy control model is used to process the vertical and horizontal feature parameters to generate vertical lifting control commands, and combined with the preset horizontal control rules to obtain horizontal displacement control commands. S424. Vertical lifting and horizontal displacement control commands are assigned to different processing threads, and multi-core processors and parallel computing frameworks are used to accelerate the generation of control commands.

[0016] According to another aspect of the present invention, an intelligent control system for obstacle avoidance of aerial work platforms is also provided, the system comprising: The sensing data acquisition and spatial construction module is used to collect sensing data of the working area of ​​the aerial work platform and construct a three-dimensional spatial map based on the sensing data. The obstacle recognition and region segmentation module is used to identify dynamic and static obstacles in the 3D spatial map and combine the topological potential field algorithm to classify the risk level of the work area. The obstacle avoidance path generation module is used to plan the initial obstacle avoidance path based on the work area division results using a random tree algorithm, and optimize the initial obstacle avoidance path in combination with the construction parameters of the aerial work platform to generate the final obstacle avoidance path. The control command and execution optimization module is used to decouple the final obstacle avoidance path into control commands for vertical lifting and horizontal displacement, execute the corresponding obstacle avoidance behavior mode based on the control commands, and monitor the execution results to optimize the control commands.

[0017] The beneficial effects of this invention are as follows: 1. By constructing a three-dimensional spatial map, this invention can not only capture the geometric features of static obstacles, but also analyze the motion characteristics of dynamic obstacles using time-series sensing data, thereby realizing the behavior modeling of dynamic obstacles, providing a comprehensive and accurate environmental cognition basis for path planning, and improving the safety and efficiency of path planning.

[0018] 2. This invention combines digital twin technology with potential field analysis algorithms to establish a motion prediction model for dynamic obstacles, and uses a topological potential field algorithm to dynamically divide the workspace into risk areas, enabling the path planner to identify and predict changes in risk areas in advance, thereby providing a valuable time window for path adjustment and ensuring safer and more efficient path decision-making.

[0019] 3. This invention achieves vertical and horizontal decoupling control of motion commands by adopting a model predictive control framework, and constructs a closed-loop optimization system by combining platform state feedback. This not only ensures the stability of the system, but also improves the flexibility and energy efficiency of obstacle avoidance behavior. It can avoid obstacles more efficiently while maintaining stability and optimize energy use efficiency. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of an intelligent obstacle avoidance control method for an aerial work platform according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an intelligent obstacle avoidance control system for aerial work platforms according to an embodiment of the present invention.

[0022] In the picture: 1. Perception data acquisition and spatial construction module; 2. Obstacle recognition and region division module; 3. Obstacle avoidance path generation module; 4. Control command and execution optimization module. Detailed Implementation

[0023] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0024] According to an embodiment of the present invention, an intelligent control method and system for obstacle avoidance of aerial work platforms are provided.

[0025] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, the intelligent control method for obstacle avoidance of an aerial work platform includes: S1. Collect sensing data of the working area of ​​the aerial work platform and construct a three-dimensional spatial map based on the sensing data.

[0026] Preferably, collecting sensing data of the working area of ​​the aerial work platform and constructing a three-dimensional spatial map based on the sensing data includes: S11. Collect sensing data of the working area of ​​the aerial work platform through multimodal sensors, and perform spatiotemporal synchronization and coordinate alignment on the collected sensing data; S12. The Kalman filter algorithm is used to fuse the sensing data after spatiotemporal synchronization and coordinate alignment, and the density-based clustering algorithm is combined to denoise the fused sensing data. S13. Divide the denoised perception data into a voxel grid, extract the centroid of each voxel as the target point, and construct a three-dimensional spatial map by stitching together the target points.

[0027] It should be further explained that, in a specific embodiment, collecting sensing data of the working area of ​​the aerial work platform and constructing a three-dimensional spatial map based on the sensing data includes: First, a LiDAR system collects 100,000 point cloud data points per second, a camera captures 30 frames of images per second, and an IMU records 100 acceleration and angular velocity sensing data points per second. Next, a GPS time synchronization algorithm is used to ensure consistent timestamps across all sensors. A calibration system aligns the coordinate systems of the LiDAR and camera to the IMU's coordinate system. Then, a Kalman filter algorithm is used to fuse the sensing data from the LiDAR, camera, and IMU to estimate the platform's precise position and trajectory. The DBSCAN algorithm is applied to denoise the LiDAR point cloud sensing data, retaining approximately 80% of the sensing data points after removing noise. Finally, the denoised point cloud sensing data is divided into a 0.1m × 0.1m × 0.1m voxel grid, and the centroid of each voxel is calculated as a target point, generating approximately 500,000 target points. By stitching together these target points, a 3D spatial map of the construction site is constructed, achieving centimeter-level accuracy.

[0028] In addition, perception data includes visual data, distance and location data, radar data, environmental information data, GPS data, and sensor fusion data.

[0029] S2. Identify dynamic and static obstacles in the 3D spatial map, and classify the risk level of the work area using the topological potential field algorithm.

[0030] S2 includes: S21. Based on a three-dimensional spatial map, use deep learning algorithms to identify and classify dynamic and static obstacles within the working area, and obtain the location information of dynamic and static obstacles.

[0031] It should be noted that static obstacles are usually fixed objects, such as buildings and equipment. These obstacles are characterized by their relatively fixed positions and minimal changes. Deep learning models obtained through pre-training can identify and classify these objects. During training, static obstacles are treated as a category, and the model learns the characteristics of static obstacles based on the spatial distribution of the point cloud.

[0032] Dynamic obstacles include moving objects such as pedestrians, vehicles, and drones. These obstacles are typically characterized by their constantly changing positions over time. Point cloud sensing data can be input into a deep learning model in time series, utilizing temporal differences to capture the features of dynamic obstacles. For dynamic obstacles, motion detection can be used to identify objects with significant positional changes between multiple time frames by comparing multiple frames of sensing data.

[0033] S22. Use digital twin technology to construct a digital twin model and combine it with potential field analysis algorithms to predict the trajectory of dynamic obstacles.

[0034] Preferably, the process of constructing a digital twin model using digital twin technology and combining it with a potential field analysis algorithm to predict the trajectory of dynamic obstacles includes: S221. Based on the location information of dynamic obstacles, a digital twin model is constructed using digital twin technology, and the motion state of the dynamic obstacles is updated through the fused real-time perception data.

[0035] S222. Perform time-frequency analysis on motion perception data in the digital twin model, extract motion cycle features, and combine graph neural networks to mine the interactive topological relationships between dynamic obstacles to generate a spatiotemporal correlation risk matrix to quantify dynamic trends and potential risks.

[0036] Preferably, time-frequency analysis is performed on the motion perception data in the digital twin model to extract motion cycle features, and graph neural networks are used to mine the interaction topological relationships between dynamic obstacles to generate a spatiotemporal correlation risk matrix, in order to quantify dynamic trends and potential risks, including: S2221. Use short-time Fourier transform to perform time-frequency analysis on motion sensing data in digital twin model, extract periodic features of motion sensing data, and obtain motion periodic features of dynamic obstacles.

[0037] S2222. Based on the motion perception data of dynamic obstacles, each dynamic obstacle is defined as a node, and the interaction relationship between obstacles is defined as an edge, thus constructing an interaction topology graph.

[0038] It should be further explained that, in this specific embodiment, it is assumed that there are four dynamic obstacles A, B, C, and D, and their position information and velocities are recorded. Based on the calculations, the following is derived: The distance between A and B is 0.5 meters, and their speed difference is 0.2 m / s, so an edge is constructed; the distance between A and C is 2 meters, and their speed difference is 1 m / s, so no edge is constructed; the distance between B and D is 1 meter, and their speed difference is 0.3 m / s, so an edge is constructed; then the interactive topology graph is: A--B, B--D.

[0039] S2223. Utilize graph neural networks to learn the interactive topology graph, identify and analyze the topological structure and dynamic interaction patterns between dynamic obstacles, and capture the spatiotemporal interaction information between dynamic obstacles.

[0040] S2224. Combining motion cycle characteristics and spatiotemporal interaction information, calculate the potential risks between dynamic obstacles, construct a spatiotemporal correlation risk matrix, and predict the dynamic changes and potential risks of obstacles by analyzing the changing trend of the spatiotemporal correlation risk matrix.

[0041] It should be further noted that, in a specific embodiment, time-frequency analysis is performed on the motion perception data in the digital twin model to extract motion cycle features, and graph neural networks are used to mine the interaction topological relationships between dynamic obstacles to generate a spatiotemporal correlation risk matrix, in order to quantify dynamic trends and potential risks, including: 1. Motion cycle feature extraction: Short-time Fourier transform is performed on motion perception data. Based on the transformation process, significant energy peaks are detected in the frequency domain. For example, the AGV1 cycle is 0.83±0.05s (corresponding to the start-stop cycle), the AGV2 cycle is 1.21±0.03s (corresponding to loading and unloading operations), and the personnel gait cycle is 1.12±0.1s.

[0042] 2. Interactive topology graph construction: An adjacency matrix is ​​established based on the minimum distance threshold between obstacles, and the edge weights are calculated using a Gaussian kernel function; ; In the formula, w ij Represents a node i and nodes j Edge weights between them; d ij Represents a node i and nodes j The distance between them; s Represents a constant; where s =0.3m, construct a time-varying undirected graph G(t)∈R 5×5 .

[0043] 3. Graph Neural Network Training: A spatiotemporal graph convolutional network is used, with 3 graph convolutional layers (64 channels per layer), a temporal convolutional kernel length of 3, and a masked cross-entropy loss function. The training set and validation set are divided in an 8:2 ratio.

[0044] 4. Risk Matrix Construction and Analysis: Integrating motion cycle characteristics (Fourier coefficients) and graph attention weights, a spatiotemporal correlation risk matrix G(t)∈R is constructed. 5×5 The results showed that when AGV1 interacts with personnel, the R value suddenly increases (the risk of collision increases), and the risk propagation exhibits an exponential decay characteristic with an impact radius of about 1.8m. Through matrix eigenvalue analysis, risk events can be predicted 0.8-1.2s in advance.

[0045] 5. Prediction Results: Comparing the actual trajectory with the predicted risk area, the system successfully issued warnings for 4 potential collisions (warning time window 1.5-2.3s), with a false positive rate controlled at 3.1%. The risk matrix visualization results show that the overlap between the high-risk area and the actual trajectory of the obstacle reaches 89.4%.

[0046] S223. Map the spatiotemporal correlation risk matrix to a potential field intensity distribution, use the fast multipole algorithm to calculate the potential field changes caused by the motion of dynamic obstacles, and establish a kinematic model of dynamic obstacles based on the potential field change results.

[0047] Preferably, the spatiotemporal correlation risk matrix is ​​mapped to a potential field intensity distribution, and the potential field changes caused by the motion of dynamic obstacles are calculated using the fast multipole algorithm. Based on the potential field change results, a kinematic model of the dynamic obstacle is established, including: S2231. Use a nonlinear mapping function to convert the spatiotemporal correlation risk matrix into a potential field intensity distribution, and assign the mapped potential field intensity values ​​to the corresponding grid nodes to construct a potential field intensity distribution map. S2232. Use an octree structure to spatially divide the potential field intensity distribution map, and perform cluster analysis based on the location information of dynamic obstacles to form a tree-like hierarchical structure. S2233. Combining the tree-like hierarchical structure with real-time motion perception data of dynamic obstacles, the fast multipole algorithm is used to calculate the potential field changes caused by the motion of dynamic obstacles and update the potential field intensity distribution. S2234. Extract the motion characteristics of the dynamic obstacle from the updated potential field intensity distribution, and establish a kinematic model of the dynamic obstacle based on the motion characteristics.

[0048] Preferably, the formula for the kinematic model is: ; In the formula, r ( t +1) indicates that the dynamic obstacle occurs in time. t +1 position;r ( t ) indicates the dynamic obstacle in time t Location; t Indicates time; v ( t ) indicates the dynamic obstacle in time t speed; Δ t Indicates the time step; F ( t ) indicates the dynamic obstacle in time t The external force caused by the potential field it experiences; m This indicates the mass of a dynamic obstacle.

[0049] It should be further explained that, in the specific embodiment, the spatiotemporal correlation risk matrix is ​​mapped to a potential field intensity distribution, and the potential field changes caused by the motion of dynamic obstacles are calculated using the fast multipole algorithm. Based on the potential field change results, a kinematic model of the dynamic obstacle is established, including: 1. Potential field intensity mapping: The risk matrix is ​​mapped using the hyperbolic tangent function. R ( t The mapping is to the potential field strength; ; In the formula, f ( x,y ) indicates at coordinate point ( x,y The potential field strength at point ( ); tanh represents the hyperbolic tangent function; R ( x,y ) indicates at coordinate point ( x,y Risk matrix value at () l Represents the scaling factor; where l =0.8; the experimentally measured potential field strength at the center of the obstacle reached 0.92 (normalized value), and the edge region showed exponential decay.

[0050] 2. Octree spatial clustering: Construct an octree structure with a depth of 5, and perform dynamic clustering based on the location of obstacles, with the maximum clustering error controlled within 0.3m; the tree node update frequency is positively correlated with the obstacle speed (update cycle is 1.2s when v=0.5m / s).

[0051] 3. Fast Multipole Update: The FMM algorithm is used to calculate the potential field change, and the expansion coefficients are set. p =6, the computational complexity is reduced from O(N²) to O(N), the time for a single update is reduced from 48ms in the traditional method to 8ms; the prediction error of potential field change is <5%.

[0052] 4. Kinematic Model Construction: Extract the motion characteristics (such as velocity, acceleration, displacement, etc.) of each dynamic obstacle from the updated potential field intensity distribution, and build a kinematic model based on these characteristics; where the change of potential field affects the acceleration or velocity of the obstacle (for example, adjusting the motion strategy by sensing the change of potential field), the change of potential field intensity is incorporated into the kinematic model as an external force influence.

[0053] S224. Combining kinematic models with long short-term memory networks, predict the trajectory of obstacles and optimize the kinematic model through real-time perception data feedback.

[0054] It should be further explained that combining kinematic models with long short-term memory networks to predict the trajectory of obstacles, and optimizing the kinematic models through real-time perception data feedback, includes: Historical sensing data (such as motion sensing data from the past 30 seconds) is used to train a long short-term memory network model, which is then combined with a kinematic model to initially model the motion of the AGV. When the AGV starts to accelerate or change direction during operation, the real-time feedback speed and acceleration sensing data is input into the LSTM model for online training. The model is adjusted according to the real-time sensing data to gradually optimize the trajectory prediction accuracy.

[0055] S23. Combining the prediction results of dynamic obstacles with the location information of static obstacles, the working area is divided into a passable area, a cautious area, and a prohibited area using the topological potential field algorithm to identify different dynamic risk levels.

[0056] It should be further explained that, in the specific embodiment, the spatiotemporal correlation risk matrix is ​​mapped to a potential field intensity distribution, and the potential field changes caused by the motion of dynamic obstacles are calculated using the fast multipole algorithm. Based on the potential field change results, a kinematic model of the dynamic obstacle is established, including: 1. Dynamic trajectory prediction: Analyze the prediction results of dynamic obstacles to obtain the trajectory of the forklift in the next 3 seconds. The average displacement error is 0.18m and the maximum error does not exceed 0.35m.

[0057] 2. Static obstacle core extraction: The point cloud sensing data is processed by the DBSCAN clustering algorithm to extract the shelf location (clustering error <0.05m) and construct the static potential field; ; In the formula, U static This represents the total potential energy of a static potential field. N Indicates the number of static obstacles; i Indicates the index value; Q i Indicates the first i The charge of each obstacle (shelf Q=5, wall Q=10); Pi Indicates the first i The location of the obstacle; X i Indicates the location of the reference point; || P i - Q i || indicates the first i The distance between the location of each obstacle and the reference point.

[0058] 3. Dynamic potential field calculation: Combining the predicted trajectory, the dynamic potential field is constructed using the velocity obstacle (VO) method; ; In the formula, U dynamic This represents the total potential energy of the dynamic potential field. M Indicates the number of dynamic obstacles; j Indicates the index value; c Indicates the attenuation coefficient ( c =0.8 is the attenuation coefficient). v j Indicates the first j The speed of the obstacle; t This represents the safety margin in terms of time (τ=1.5s is the time margin). d min represents the minimum safe distance from the obstacle. d The potential field strength changes abruptly when min=0.5m.

[0059] 4. Risk Area Division: Overall Potential Field U = U static + U dynamic Set thresholds: for the passable area, U < 0.3; for the cautious area, 0.3 ≤ U < 0.7; for the prohibited area, U ≥ 0.7.

[0060] S3. Based on the work area division results, the initial obstacle avoidance path is planned using the random tree algorithm. The initial obstacle avoidance path is then optimized by combining the construction parameters of the aerial work platform to generate the final obstacle avoidance path.

[0061] Preferably, based on the work area division results, an initial obstacle avoidance path is planned using a random tree algorithm. This initial obstacle avoidance path is then optimized by combining it with the construction parameters of the aerial work platform, generating the final obstacle avoidance path, which includes: S31. Based on the region division results, construct a three-dimensional grid map and use a Gaussian mixture model to fit the obstacle distribution boundary to generate obstacle avoidance constraint boundaries.

[0062] It should be noted that the formula for the Gaussian mixture model is: ; In the formula, p ( x t ) indicates the position x The probability density function at that location; K This indicates the number of Gaussian components in a Gaussian mixture model. k Indicates the index value; π k ( t ) indicates the first k The Gaussian component in time t The weights; N Indicates a Gaussian distribution; x t Indicates time t Location; m k ( t ) indicates the first k The Gaussian component in time t The mean vector; ∑ k ( t ) indicates the first k The Gaussian component in time t The covariance matrix.

[0063] S32. Based on the obstacle avoidance constraint boundary, use the target bias-based random tree algorithm to perform path search and generate the initial obstacle avoidance path.

[0064] S33. Obtain the construction parameters of the aerial work platform, define an optimization target set including path length, minimum safe distance, curvature smoothness and energy consumption coefficient, and use the entropy weight algorithm to dynamically adjust the weight of the optimization target according to the control task requirements.

[0065] S34. Based on the weight adjustment results, the initial obstacle avoidance path is adjusted using the particle swarm optimization algorithm to generate the final obstacle avoidance path.

[0066] It should be further explained that, in the specific embodiment, based on the work area division results, an initial obstacle avoidance path is planned using a random tree algorithm. This initial obstacle avoidance path is then optimized by combining it with the construction parameters of the aerial work platform, generating the final obstacle avoidance path, which includes: 1. 3D raster modeling and Gaussian mixture model fitting: Construct a voxel map and use GMM to fit obstacle boundaries.

[0067] 2. Path search based on target bias: Set the target bias probability and step size to generate the initial path length.

[0068] 3. Multi-objective optimization modeling: Obtain UAV parameters; for example, arm length 0.8m → safety radius constraint 1.2m, battery capacity 6S 12000mAh → energy consumption coefficient 0.05Wh / m; define the objective function; ; In the formula, J This represents the value of the integrated objective function; L Indicates the path length; d min represents the minimum safe distance from the obstacle; k Indicates path curvature; E Indicates estimated energy consumption; w 1 、w 2 、w 3 、w 4 represents the weighting coefficient.

[0069] 4. Dynamic weight adjustment and particle swarm optimization: The entropy weight method is used to calculate the weights; ; In the formula, w Indicates weight; H This represents the entropy value.

[0070] For example, the particle swarm optimization parameters are: 50 particles, 100 iterations, and an inertia weight of 0.729. The optimized paths are shown in Table 1.

[0071] Table 1 Path Comparison Perception Data Table S4. Decouple the final obstacle avoidance path into control commands for vertical lifting and horizontal displacement, execute the corresponding obstacle avoidance behavior mode based on the control commands, and monitor the execution results to optimize the control commands.

[0072] Preferably, the final obstacle avoidance path is decoupled into control commands for vertical lifting and horizontal displacement. Based on these control commands, corresponding obstacle avoidance behavior modes are executed, and the execution results are monitored to optimize the control commands. S41. Based on the preset spatial coordinates, the final obstacle avoidance path is decoupled into a sequence of vertical and horizontal coordinates, and a polynomial fitting technique is used to generate the motion profile of the aerial work platform.

[0073] It should be noted that the steps for generating the motion profile of the aerial work platform using polynomial fitting technology are as follows: 1. Polynomial fitting: In the horizontal direction, a 5th-order polynomial is used for fitting; ; ; Vertical direction: 7th order polynomial fitting was used; ; In the formula, x ( t ) represents the horizontal direction over time t Changing coordinates; y ( t ) represents the horizontal direction over time t Changing coordinates; z ( t ) indicates the change in the vertical direction over time t Changing coordinates; a k Indicates the level coefficient; b k Indicates the level coefficient; c k Indicates the vertical coefficient; k Indicates the index value.

[0074] 2. Motion Profile Generation: Calculate motion parameters in each direction, where velocity is... v ( t )= dx / dt ,d y / dt , dz / dt acceleration is accelerometer is The constraints include a maximum velocity of 3 m / s and a maximum acceleration of 2 m / s². 2 Continuous jerk (jerk < 0.5 m / s²) 3 ).

[0075] S42. Analyze the motion profile, use fuzzy control algorithm to generate control commands for vertical lifting and horizontal displacement, and accelerate the generation of control commands through a parallel computing framework.

[0076] Preferably, the motion profile is analyzed, and control commands for vertical lifting and horizontal displacement are generated using a fuzzy control algorithm. The generation of these control commands is accelerated using a parallel computing framework, including: S421. Extract vertical and horizontal feature parameters from the generated motion profile to establish a fuzzy input space. The feature parameters include height deviation, velocity and curvature error. S422. Combining fuzzy input space and fuzzy control algorithm, a fuzzy control model is constructed, which includes an input layer, a membership degree calculation layer and a fuzzy inference layer. A rule reduction layer is introduced to optimize the fuzzy control model. S423. The optimized fuzzy control model is used to process the vertical and horizontal feature parameters to generate vertical lifting control commands, and combined with the preset horizontal control rules to obtain horizontal displacement control commands. S424. Vertical lifting and horizontal displacement control commands are assigned to different processing threads, and multi-core processors and parallel computing frameworks are used to accelerate the generation of control commands.

[0077] It should be further explained that, in the specific embodiment, analyzing the motion profile, using a fuzzy control algorithm to generate control commands for vertical lifting and horizontal displacement, and accelerating the generation of control commands through a parallel computing framework includes: 1. Feature parameter extraction and fuzzy space establishment: Extracting motion profile features: height deviation Δ h (Actual height - Reference height); Speed ​​deviation Δ v (Actual speed - preset speed); curvature error k e (Actual curvature - Desired curvature); Establish a fuzzy input space: Δ h Domain of discourse, [-3,3]m; Δ v Domain of discourse, [-1,1] m / s; k e The domain is [-0.2, 0.2]; trigonometric membership functions are used with an overlap rate of 0.3.

[0078] 2. Fuzzy control model construction and optimization: 49 initial rules were designed (7 input combinations × 7 output combinations); a rule reduction layer was introduced, redundant rules were removed based on PCA analysis, and the rule base was optimized using a genetic algorithm. After optimization, the number of rules was reduced to 19, and the computational load was reduced by 58%.

[0079] 3. Control command generation and parallel processing: Vertical control commands; ; In the formula, u z Indicates control commands in the vertical direction; K p Indicates proportional gain; Δ h Indicates height deviation; K d This represents the differential gain.

[0080] Vertical control commands: ; In the formula, u xy Indicates control commands in the horizontal direction; PID This represents a proportional-integral-derivative controller; Δ p Indicates height deviation; Δ v Indicates speed deviation; f Δ pdt This represents a portion of the integral term, indicating the integral of the position deviation over time.

[0081] 4. Quantitative Evaluation Indicators: Define the comprehensive control performance index: Results showed: average height tracking error <0.4m; speed response latency <80ms; CPU+GPU heterogeneous computing speedup ratio 4.2x.

[0082] S43. Based on the generated control commands, select and execute the corresponding obstacle avoidance behavior mode, which includes the emergency braking model and the dynamic detour behavior mode.

[0083] S44. Monitor the execution effect of the selected obstacle avoidance behavior mode in real time, and optimize control commands based on the monitoring results.

[0084] It should be further explained that, in the specific embodiment, the final obstacle avoidance path is decoupled into control commands for vertical lifting and horizontal displacement. Based on these control commands, corresponding obstacle avoidance behavior modes are executed, and the execution results are monitored to optimize the control commands. This includes: 1. Motion profile generation: Decompose the final obstacle avoidance path into a vertical coordinate sequence ( z i ) and horizontal coordinate sequence ( x i , y i The vertical direction (7th order polynomial, fitting error <0.15m) and the horizontal direction (5th order polynomial, fitting error <0.2m) are subjected to polynomial fitting to generate a profile.

[0085] 2. Fuzzy control instruction generation: Feature parameters, height deviation Δ h (Domain of discourse [-2,2]m), Horizontal position deviation Δ p (Domain of discourse [-3,3]m), velocity deviation Δ v (Domain of discourse [-1,1] m / s); Fuzzy rule base, 3-input × 3-output system, 27 rules in total, using triangular membership functions, overlap rate 0.4; Parallel computing acceleration: CUDA kernel function to implement fuzzy inference; GPU speedup 3.8x.

[0086] 3. Obstacle avoidance behavior pattern execution: |Δ h |>1.2m→Emergency braking (decelerate to 0.5m / s);|Δ p |>2m→Dynamic detour (triggers local replanning); others→normal tracking (PID control); behavior mode switching delay <50ms (measured value).

[0087] 4. Real-time monitoring and optimization: Monitoring indicators: obstacle avoidance success rate (number of successful avoidances / total number of avoidances); control command response time; trajectory tracking error; optimization strategy: update fuzzy rule weights every 10 control cycles; use reinforcement learning to dynamically adjust PID parameters.

[0088] like Figure 2 As shown, according to another embodiment of the present invention, an intelligent obstacle avoidance control system for aerial work platforms is also provided, the system comprising: The sensor data acquisition and spatial construction module 1, obstacle recognition and region division module 2, obstacle avoidance path generation module 3, and control command and execution optimization module 4 are connected in sequence.

[0089] The sensing data acquisition and spatial construction module 1 is used to collect sensing data of the working area of ​​the aerial work platform and construct a three-dimensional spatial map based on the sensing data. The obstacle recognition and region division module 2 is used to identify dynamic and static obstacles in the three-dimensional spatial map and to classify the risk level of the work area by combining the topological potential field algorithm. The obstacle avoidance path generation module 3 is used to plan the initial obstacle avoidance path based on the work area division results using a random tree algorithm, optimize the initial obstacle avoidance path in combination with the construction parameters of the aerial work platform, and generate the final obstacle avoidance path. The control command and execution optimization module 4 is used to decouple the final obstacle avoidance path into control commands for vertical lifting and horizontal displacement, execute the corresponding obstacle avoidance behavior mode based on the control commands, and monitor the execution results to optimize the control commands.

[0090] In summary, by utilizing the technical solutions described above in this invention, a three-dimensional spatial map can be constructed, capturing not only the geometric features of static obstacles but also analyzing the motion characteristics of dynamic obstacles using time-series sensing data. This enables behavioral modeling of dynamic obstacles, providing a comprehensive and accurate environmental awareness foundation for path planning and significantly improving its safety and efficiency. By combining digital twin technology with potential field analysis algorithms, a motion prediction model for dynamic obstacles is established. Furthermore, a topological potential field algorithm is used to dynamically divide the workspace into risk areas, allowing the path planner to identify and predict changes in risk areas in advance. This provides a valuable time window for path adjustments, ensuring safer and more efficient path decisions. By employing a model predictive control framework to achieve vertical and horizontal decoupling of motion commands and combining this with platform state feedback to construct a closed-loop optimization system, system stability is ensured, and the flexibility and energy efficiency of obstacle avoidance behavior are significantly improved. This allows for more agile and efficient obstacle avoidance while maintaining stability, optimizing energy usage efficiency.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent obstacle avoidance control of aerial work platforms, characterized in that, The method includes: S1. Collect sensing data of the working area of ​​the aerial work platform and construct a three-dimensional spatial map based on the sensing data; S2. Identify dynamic and static obstacles in the 3D spatial map, and classify the risk level of the work area using the topological potential field algorithm; S3. Based on the work area division results, the initial obstacle avoidance path is planned using the random tree algorithm. The initial obstacle avoidance path is then optimized by combining the construction parameters of the aerial work platform to generate the final obstacle avoidance path. S4. Decouple the final obstacle avoidance path into control commands for vertical lifting and horizontal displacement, execute the corresponding obstacle avoidance behavior mode based on the control commands, and monitor the execution results to optimize the control commands. Wherein, S2 includes: S21. Based on a three-dimensional spatial map, use deep learning algorithms to identify and classify dynamic and static obstacles within the working area, and obtain the location information of dynamic and static obstacles; S22. Use digital twin technology to construct a digital twin model and combine it with potential field analysis algorithm to predict the motion trajectory of dynamic obstacles; S23. Combining the prediction results of dynamic obstacles with the location information of static obstacles, the working area is divided into a passable area, a cautious area, and a prohibited area using the topological potential field algorithm to identify different risk levels.

2. The intelligent control method for obstacle avoidance of an aerial work platform according to claim 1, characterized in that, The process of collecting sensor data of the working area of ​​the aerial work platform and constructing a three-dimensional spatial map based on the sensor data includes: S11. Collect sensing data of the working area of ​​the aerial work platform through multimodal sensors, and perform spatiotemporal synchronization and coordinate alignment on the collected sensing data; S12. The Kalman filter algorithm is used to fuse the sensing data after spatiotemporal synchronization and coordinate alignment, and the density-based clustering algorithm is combined to denoise the fused sensing data. S13. Divide the denoised perception data into a voxel grid, extract the centroid of each voxel as the target point, and construct a three-dimensional spatial map by stitching together the target points.

3. The intelligent control method for obstacle avoidance of an aerial work platform according to claim 1, characterized in that, The method of constructing a digital twin model using digital twin technology and combining it with a potential field analysis algorithm to predict the trajectory of dynamic obstacles includes: S221. Based on the location information of dynamic obstacles, a digital twin model is constructed using digital twin technology, and the motion state of the dynamic obstacles is updated through the fused real-time perception data. S222. Perform time-frequency analysis on motion perception data in the digital twin model, extract motion cycle features, and combine graph neural networks to mine the interactive topological relationships between dynamic obstacles to generate a spatiotemporal correlation risk matrix to quantify dynamic trends and potential risks. S223. Map the spatiotemporal correlation risk matrix to a potential field intensity distribution, use the fast multipole algorithm to calculate the potential field changes caused by the motion of dynamic obstacles, and establish a kinematic model of dynamic obstacles based on the potential field change results. S224. Combining kinematic models with long short-term memory networks, predict the trajectory of obstacles and optimize the kinematic model through real-time perception data feedback.

4. The intelligent control method for obstacle avoidance of an aerial work platform according to claim 3, characterized in that, The process of performing time-frequency analysis on motion perception data in the digital twin model, extracting motion cycle features, and combining graph neural networks to mine the interaction topological relationships between dynamic obstacles to generate a spatiotemporal correlation risk matrix to quantify dynamic trends and potential risks includes: S2221. Use short-time Fourier transform to perform time-frequency analysis on motion sensing data in digital twin model, extract periodic features of motion sensing data, and obtain motion periodic features of dynamic obstacles. S2222. Based on the motion perception data of dynamic obstacles, each dynamic obstacle is defined as a node, and the interaction relationship between obstacles is defined as an edge, thus constructing an interaction topology graph; S2223. Use graph neural networks to learn the interaction topology graph, identify and analyze the topological structure and dynamic interaction patterns between dynamic obstacles, and capture the spatiotemporal interaction information between dynamic obstacles. S2224. Combining motion cycle characteristics and spatiotemporal interaction information, calculate the potential risks between dynamic obstacles, construct a spatiotemporal correlation risk matrix, and predict the dynamic changes and potential risks of obstacles by analyzing the changing trend of the spatiotemporal correlation risk matrix.

5. The intelligent control method for obstacle avoidance of an aerial work platform according to claim 4, characterized in that, The process of mapping the spatiotemporal correlation risk matrix to a potential field intensity distribution, calculating the potential field changes caused by the motion of dynamic obstacles using the fast multipole algorithm, and establishing a kinematic model of the dynamic obstacles based on the potential field change results includes: S2231. Use a nonlinear mapping function to convert the spatiotemporal correlation risk matrix into a potential field intensity distribution, and assign the mapped potential field intensity values ​​to the corresponding grid nodes to construct a potential field intensity distribution map. S2232. Use an octree structure to spatially divide the potential field intensity distribution map, and perform cluster analysis based on the location information of dynamic obstacles to form a tree-like hierarchical structure. S2233. Combining the tree-like hierarchical structure with real-time motion perception data of dynamic obstacles, the fast multipole algorithm is used to calculate the potential field changes caused by the motion of dynamic obstacles and update the potential field intensity distribution. S2234. Extract the motion characteristics of the dynamic obstacle from the updated potential field intensity distribution, and establish a kinematic model of the dynamic obstacle based on the motion characteristics.

6. The intelligent obstacle avoidance control method for aerial work platforms according to claim 5, characterized in that, The formula for the kinematic model is: ； In the formula, r ( t +1) indicates that the dynamic obstacle occurs in time. t +1 position; r ( t ) indicates the dynamic obstacle in time t Location; t Indicates time; v ( t ) indicates the dynamic obstacle in time t speed; Δ t Indicates the time step; F ( t ) indicates the dynamic obstacle in time t The external force caused by the potential field it experiences; m This indicates the mass of a dynamic obstacle.

7. The intelligent control method for obstacle avoidance of an aerial work platform according to claim 1, characterized in that, The process of planning an initial obstacle avoidance path based on the work area division results using a random tree algorithm, optimizing the initial obstacle avoidance path in conjunction with the construction parameters of the aerial work platform, and generating the final obstacle avoidance path includes: S31. Based on the region division results, construct a three-dimensional grid map and use a Gaussian mixture model to fit the obstacle distribution boundary to generate obstacle avoidance constraint boundary. S32. Based on the obstacle avoidance constraint boundary, use the target bias-based random tree algorithm to perform path search and generate the initial obstacle avoidance path; S33. Obtain the construction parameters of the aerial work platform, define an optimization target set including path length, minimum safe distance, curvature smoothness and energy consumption coefficient, and use the entropy weight algorithm to dynamically adjust the weight of the optimization target according to the control task requirements. S34. Based on the weight adjustment results, the initial obstacle avoidance path is adjusted using the particle swarm optimization algorithm to generate the final obstacle avoidance path.

8. The intelligent control method for obstacle avoidance of an aerial work platform according to claim 1, characterized in that, The process of decoupling the final obstacle avoidance path into control commands for vertical lifting and horizontal displacement, executing corresponding obstacle avoidance behavior modes based on the control commands, and monitoring the execution results to optimize the control commands includes: S41. Based on the preset spatial coordinates, the final obstacle avoidance path is decoupled into a sequence of vertical and horizontal coordinates, and a polynomial fitting technique is used to generate the motion profile of the aerial work platform. S42. Analyze the motion profile, use fuzzy control algorithm to generate control commands for vertical lifting and horizontal displacement, and accelerate the generation of control commands through a parallel computing framework; S43. Select and execute the corresponding obstacle avoidance behavior mode according to the generated control command. The obstacle avoidance behavior mode includes an emergency braking model and a dynamic detour behavior mode. S44. Monitor the execution effect of the selected obstacle avoidance behavior mode in real time, and optimize control commands based on the monitoring results.

9. The intelligent control method for obstacle avoidance of an aerial work platform according to claim 8, characterized in that, The analysis of the motion profile utilizes a fuzzy control algorithm to generate control commands for vertical lifting and horizontal displacement, and accelerates the generation of control commands through a parallel computing framework, including: S421. Extract vertical and horizontal feature parameters from the generated motion profile to establish a fuzzy input space. The feature parameters include height deviation, velocity, and curvature error. S422. Combining fuzzy input space and fuzzy control algorithm, a fuzzy control model is constructed, which includes an input layer, a membership degree calculation layer and a fuzzy inference layer. A rule reduction layer is introduced to optimize the fuzzy control model. S423. The optimized fuzzy control model is used to process the vertical and horizontal feature parameters to generate vertical lifting control commands, and combined with the preset horizontal control rules to obtain horizontal displacement control commands. S424. Vertical lifting and horizontal displacement control commands are assigned to different processing threads, and multi-core processors and parallel computing frameworks are used to accelerate the generation of control commands.

10. An intelligent obstacle avoidance control system for aerial work platforms, used to implement the intelligent obstacle avoidance control method for aerial work platforms as described in any one of claims 1-9, characterized in that, The system includes: The sensing data acquisition and spatial construction module is used to collect sensing data of the working area of ​​the aerial work platform and construct a three-dimensional spatial map based on the sensing data. The obstacle recognition and region segmentation module is used to identify dynamic and static obstacles in the 3D spatial map and combine the topological potential field algorithm to classify the risk level of the work area. The obstacle avoidance path generation module is used to plan the initial obstacle avoidance path based on the work area division results using a random tree algorithm, and optimize the initial obstacle avoidance path in combination with the construction parameters of the aerial work platform to generate the final obstacle avoidance path. The control command and execution optimization module is used to decouple the final obstacle avoidance path into control commands for vertical lifting and horizontal displacement, execute the corresponding obstacle avoidance behavior mode based on the control commands, and monitor the execution results to optimize the control commands.

Images