A path planning method for large bulk material equipment based on risk boundary

Abstract

The present invention discloses a path planning method for large-scale bulk material equipment based on risk boundaries, which involves the fields of machine learning, path planning, etc. First, deterministic point clouds are used to obtain obstacles in the environment, and the boundaries of the safe area under a given risk tolerance are calculated to define the risk boundaries of the obstacles, and a neural network sampler is trained through expert demonstration learning decision strategies. Subsequently, a neural cost estimator is constructed to select the best information state from the sample, and a neural network is used to recursively advance in both directions toward the starting point and the target, and then the collision risk of the intermediate connection is verified based on the sum of squares optimization. Compared with other methods, the present invention can not only be applied to path planning of large-scale bulk material equipment, but can also be extended to other fields and scenarios, such as unmanned vehicles, drone navigation, environmental monitoring, etc.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning, path planning, and in particular to a path planning method for large-scale bulk material equipment based on risk boundaries. Background Art

[0002] Large bulk material handling equipment is a heavy-duty bulk material handling and transportation machinery system with continuous processing capabilities. It is widely used in port loading and unloading, power plants, mineral resource development, and equipment manufacturing. With the development of intelligent transformation and digital transformation in various industries in my country, smart ports, smart power plants, smart mines, and smart factories have become an inevitable trend in the development of the industrial field. In this context, unmanned operation of large bulk material handling equipment is an important part of improving the automation of production processes and promoting intelligent transformation and digital transformation. In the complex environment of smart ports, which involve dense multi-factors such as people, vehicles, objects, and ships, how to enable the system to accurately locate and plan the optimal safe path during cantilever operations to achieve safe and efficient automatic stacking and reclaiming operations has become a core challenge in the development of unmanned systems.

[0003] Currently, path planning solutions for cantilevered arms of large bulk material handling equipment remain immature, with manual experience and semi-automated approaches often used to locate reclaim points and plan reclaim paths. Some researchers have simplified the bucket wheel machine's operating environment model into a convex shape and assumed that path uncertainty follows a Gaussian distribution to achieve autonomous path planning. This approach, by simplifying the complexity of the environment, makes the path planning problem more tractable. However, this simplification may overlook the non-convexity and non-Gaussian properties found in real-world scenarios such as ports and mines, increasing the likelihood of the system encountering safety risks such as collisions during operation. These methods have achieved some success in closed, single-site environments. However, in complex and dynamic real-world scenarios such as ports and mines, ignoring these non-convexity and non-Gaussian properties can lead to safety risks such as collisions during operation.

[0004] To address these challenges and practical needs, this paper proposes a risk-bounded path planning method for large-scale bulk material handling equipment. This method aims to plan near-optimal paths with acceptable collision probabilities and enables large-scale bulk material handling equipment to autonomously learn risk-bounded path planning. This method first uses deterministic point cloud maps to identify potential hazards such as mobile machinery, unstable ground, and dynamically changing material storage areas in ports, mines, or other bulk material handling sites. Risk boundaries are then defined to clarify the risk levels and acceptable operating ranges for different areas. A neural network sampler (NNS) is trained through expert demonstrations to predict the state distribution in the latent subspace of large-scale bulk material handling equipment, such as material flow and equipment movement. Next, a neural cost estimator (NCE) is constructed to filter the optimal information state from the sample set. This helps large-scale bulk material handling equipment select the lowest-cost and safest path among multiple possible paths, while also considering operational efficiency. Utilizing a neural network recursive algorithm, progressing bidirectionally from the starting point to the destination, and using the sum-of-squares optimization (SOS) method to verify collision risk along the intermediate path, this method comprehensively considers path safety and efficiency, particularly in the face of a constantly changing operating environment. Through this process, the present invention can effectively reduce operational risks, optimize equipment maintenance strategies, and enhance environmental performance, thereby minimizing various losses while ensuring production safety. Furthermore, this method performs path planning in uncertain environments and provides effective path solutions. Summary of the Invention

[0005] The present invention proposes a path planning method for large-scale bulk material equipment based on risk boundaries, aiming to efficiently plan a path with guaranteed risk tolerance for large-scale bulk material equipment in an uncertain environment.

[0006] The specific implementation steps are as follows:

[0007] Step 1: In an uncertain environment, first use Gazebo to build a bulk material equipment environment, and then define the path planning problem, including the safe space, the uncertain dangerous environment, the safe initial state, the target state, and the acceptable risk level.

[0008] First, suppose , Represents the working space for the operation of large bulk material equipment. , represents the static risk area in the workspace. The danger zone is described as a polynomial on the coordinate position as :

[0009] ,

[0010] in, is a polynomial function of obstacles and can describe risk areas with uncertain size, location or geometry.

[0011] set up It is the configuration space of bulk material equipment, where There are obstacles and risk bounded Configuration space. Let and As the safe starting configuration and target configuration respectively. The goal of bulk equipment is to find the following risk-bounded path:

[0012] ,

[0013] in Indicates inclusion Configuration and connection in and risky paths.

[0014] Based on the above assumptions, the location of the path Given The risk-bounded optimal path planning problem is defined as:

[0015] ,

[0016] Constraints ,

[0017] The cost function is the path length measured by Euclidean distance , Represents the set of all feasible paths.

[0018] The probability constraint expressing the collision risk is:

[0019] ,

[0020] Step 2: Risk Contour Mapping: Create a risk contour map to describe the probabilistic information of the uncertain hazardous environment and divide the environment into safe, dangerous, and risky areas. Use a neural network to encode the point cloud map and convert it into a potential embedding representation, allowing the neural network to learn and understand the environment.

[0021] Will Risk Profile Defined as the following set of states:

[0022] ,

[0023] The deterministic constraints on the safe state set are approximately:

[0024] ,

[0025] in Indicates that obstacles are The mean of the locations, Indicates that obstacles are The variance under position,

[0026] gather is the original risk profile An inner approximation based on rational polynomials. In the risk-bounded state set The resulting path consisting of these states Guaranteed not more than The risk profile is as follows:

[0027] ,

[0028] , is the target state,

[0029] ,

[0030] Risk areas and safe areas are defined by risk contours in the point cloud. The risk area (larger blue sphere) represents the space where the probability of collision with the uncertain hazardous environment is greater than a certain threshold, while the collision probability in the safe area (outside the area) does not exceed the threshold.

[0031] Step 3: Construct and train a neural network. First, use RRT-SOS in the environment around the bulk material equipment to obtain a near-optimal solution given a large number of start and target state pairs. Based on this, two deep neural networks, NNS and NCE, are constructed to simulate the behavior of RRT-SOS and generate critical states with information. Through training, NNS is able to extract potential embedding features from the point cloud map and predict the state distribution in the latent subspace. NNS is trained using expert demonstration data to generate informative states to guide path planning. NCE is used to evaluate the cost from the current state to the target state and select the best state from it. The inputs of NNS and NCE are point cloud maps, respectively. , current status ,Target and specified risk tolerance NNS outputs the next promising state, and NCE outputs the cost estimate from the current state to the goal. The neural network is trained using mean squared error loss as the loss function, and point cloud data is used to represent the three-dimensional environment to provide input information for the neural network.

[0032] Step 3.1: The NNS consists of a barrier encoder network (PCnet) and an inference network (Inet). The PCnet has four key building blocks for extracting the latent embedding Z: a ​​Transformer Network (T-Net), a CNN, a max pooling layer, and a Multilayer Perceptron (MLP). The T-Net is a mini-network that performs a set of convolutional and fully connected operations on the raw sample data.

[0033] Step 3.2: NCE consists of the same encoder and the encoder for cost prediction. Considering the randomness of the neural network system, let the neural network system process a batch of current states , and generate several promising states Then, the NCE forecast reaches The cost required to improve the online planning performance target. Note that the tree will be pruned by eliminating redundant nodes. This method attempts to connect two nodes that are far apart. If a node has a high target cost value, there may be some nodes in the area between the target and itself, making the long-distance connection likely to fail. The node with the lowest cost is found in the iteration, the critical state is selected, and the planning time is saved.

[0034] Step 4: A bidirectional search strategy is employed, combining NNS, NCE, and a risk assessor (RA) to create a bidirectional search path planning method, known as 3D NR-RRT. Starting from the starting point and the goal, the method gradually advances toward the intermediate region. This method is used to achieve collision avoidance and find a risk-bounded path in an uncertain environment. This algorithm uses NNS and NCE to bidirectionally search for a coarse global path. NNS expands the search tree, while NCE helps find a low-cost tree.

[0035] Step 4.1: RA is used to verify the path The collision risk of a given configuration node Location , if all risk profiles If the inequality in does not hold, the nodes in the risk area are deleted. Then, the collision risk of each edge between two adjacent nodes is verified by the SOS condition. There are three steps to verify the SOS condition using the spotless toolbox:

[0036] Step 1: Export linear trajectory , given the positions of two nodes as the endpoints of the line segment;

[0037] Step 2: List the time Semi-algebraic constraints , and created multiple multiplier variables ,in is a free variable, Polynomial coefficients for constructing and verifying path security;

[0038] Step 3: Mosek solver is used to minimize the objective function , which helps determine whether the state of the expression is primal feasible and dual feasible. If it does not exist, the line segment does not have the SOS form and the probability of a collision along it is greater than .

[0039] Step 4.2: Use the 3D NR-RRT algorithm with the input being the starting and target configuration pairs { , }, risk tolerance , and risk profile and its corresponding point cloud First, use PCnet to Encoded to In an environment with Initialized to empty. Create a forward node set containing the start state and the set of backward nodes with the target state . and The update process is the same. For each iteration, a batch of samples is generated by NNS. is the local starting point, NCE prediction Choose the one with the lowest cost. and add it to If and The connection of the top node in the equation satisfies the SOS condition, then it is considered that the node containing and Rough global path of the sample If not, an empty path is returned.

[0040] In getting Then, Lazy States Contraction (LSC) is used to remove The redundant nodes in the The non-adjacent nodes of the two nodes are used to form a shortcut segment. If the shortcut segment is within the safe area, the nodes between the two ends are removed. The remaining nodes may become part of the final solution. Then, the RA method is used to evaluate the path. And get the success mark , risk edge index , and the global path of the security node .

[0041] If the path succeeds and the risk is bounded, return the path Otherwise, use Replan to replan the risk segment and use LSC and RA methods to update the path. If a feasible path is found, return the updated path. .

[0042] This method designs a path planning method for large-scale bulk material equipment based on risk boundaries, which is used to solve the path planning of large-scale bulk material equipment in uncertain environments. Compared with the existing technology, the technical solution of this invention has the following beneficial effects:

[0043] 1. The present invention builds a simulation environment for large-scale bulk material equipment, obtains obstacles in the environment through point cloud data, calculates the boundaries of the safe area under a given risk tolerance, and ultimately defines the risk boundaries of the obstacles. This allows for reasonable path planning while ensuring safety, overcoming the limitations of traditional methods. This enables safe, collision-free, autonomous operation of large-scale bulk material equipment in uncertain environments, reducing the risk of safety accidents.

[0044] 2. By learning decision-making strategies through expert demonstrations using deep neural networks, it can imitate the choices and response behaviors of experts in different environments and situations, guiding path planning in new and unknown situations. The 3D NR-RRT algorithm can handle obstacles of various shapes and sizes, improving the efficiency and quality of path planning in complex and dynamic environments, enhancing generalization, robustness and adaptability, and thus better adapting to complex and changing practical application scenarios, especially in unpredictable and highly variable environments such as ports and mines.

[0045] 3. A risk-boundary-based path planning method for large-scale bulk material equipment utilizes machine learning and neural networks to plan safe and efficient paths. This method is able to handle environmental uncertainty and complexity and can be extended to applications in areas such as unmanned vehicles, robotic navigation, and abnormal environment monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is an overall flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] See attached Figure 1This embodiment provides a path planning method for large-scale bulk material equipment facing risk boundaries in an uncertain environment. Taking a bucket wheel stacker reclaimer robot arm as an example, the method specifically includes the following processes: building a bucket wheel stacker reclaimer robot arm in a dock environment, creating a risk contour map to describe the probabilistic information of the uncertain dangerous environment, constructing a neural network sampler and a neural cost estimator to simulate RRT-SOS and generate informed critical states, and using 3D NR-RRT to find a risk-bounded path.

[0049] The specific facts are as follows:

[0050] Implementation step 1: In a deep learning system, the interaction between the agent and the environment constitutes the learning process. In this embodiment, the environment is set as a medium-to-large port, and the agent is the bucket wheel stacker cantilever. The environment is determined and the action space of the bucket wheel stacker cantilever is defined as follows.

[0051] Step 1.1: Build the bucket wheel stacker cantilever environment and set , Represents the working space of bucket wheel stacker reclaimer boom operation. , represents the static risk area in the workspace. The danger zone is described as a polynomial on the coordinate position as :

[0052] ,

[0053] in, is a given polynomial that can describe risk areas with uncertain size, location or geometry.

[0054] Step 1.2: Define the path planning problem. is the cantilever configuration space of the bucket wheel stacker reclaimer, where There are obstacles and risk bounded Configuration space. Let and They are respectively used as safe starting configuration and target configuration. The goal is to find the following risk-bounded path:

[0055] ,

[0056] in Indicates inclusion Configuration and connection in and According to the above assumptions, the location of the path Given The risk-bounded optimal path planning problem is defined as:

[0057] ,

[0058] Constraints

[0059] The cost function is the path length measured by Euclidean distance , Represents the set of all feasible paths.

[0060] The probability constraint expressing the collision risk is:

[0061] ,

[0062] Implementation Step 2: Create a risk contour map to describe the probabilistic information of the port's uncertain and hazardous environment, dividing the environment into safe zones, dangerous zones, and risk zones. The safe zone represents areas where bucket wheel excavators can safely operate, the dangerous zone is prohibited, and the risk zone is where entry may be necessary despite a certain safety risk. A neural network is used to encode the point cloud map and convert it into a latent embedding representation, enabling the neural network to learn and understand the environment.

[0063] Will Risk Profile Defined as the following set of states:

[0064] ,

[0065] Then, the deterministic constraint on the set of safe states is approximated as:

[0066] ,

[0067] in Indicates that obstacles are The mean of the locations, Indicates that obstacles are The variance under position,

[0068] gather is the original risk profile An inner approximation based on rational polynomials. In the risk-bounded state set The resulting path consisting of these states Guaranteed not more than The risk profile is as follows:

[0069] ,

[0070] , is the target state,

[0071] ,

[0072] Risk areas and safe areas are defined by risk contours in the point cloud. The risk area (larger blue sphere) represents the space where the probability of collision with the uncertain hazardous environment is greater than a certain threshold, while the collision probability in the safe area (outside the area) does not exceed the threshold.

[0073] Implementation step 3: First, use RRT-SOS in the environment around the bulk material equipment to obtain a near-optimal solution given a large number of start and target state pairs. Based on this, two deep neural networks, namely NNS and NCE, are constructed to simulate the behavior of RRT-SOS and generate critical states with information. Through training, NNS is able to extract potential embedding features from the point cloud map and predict the state distribution in the latent subspace. NNS is trained using expert demonstration data to generate informative states to guide path planning. NCE is used to evaluate the cost from the current state to the target state and select the best state from it. The inputs of NNS and NCE are point cloud maps, respectively. , current status ,Target and specified risk tolerance NNS outputs the next promising state, and NCE outputs the cost estimate from the current state to the goal. The neural network is trained using mean squared error loss as the loss function, and point cloud data is used to represent the three-dimensional environment to provide input information for the neural network.

[0074] Step 3.1: The NNS consists of a barrier encoder network (PCnet) and an inference network (Inet). The PCnet has four key building blocks for extracting the latent embedding Z: a ​​transformer network (T-Net), a CNN, a max pooling layer, and a multi-layer perceptron (MLP). The T-Net is a mini-network that performs a set of convolutional and fully connected operations on the raw sample data.

[0075] Step 3.2: NCE consists of the same encoder and the encoder for cost prediction. Considering the randomness of the neural network system, let the neural network system process a batch of current states , and generate several promising states Then, the NCE forecast reaches Improve the cost of online planning performance goals. The tree will be pruned by eliminating redundant nodes. This method attempts to connect two nodes that are far apart. If the target cost value of a node is high, there may be some nodes in the area between the target and itself, making the long-distance connection likely to fail. The node with the lowest cost is found in the iteration, the critical state is selected, and the planning time is saved.

[0076] Implementation Step 4: A bidirectional search strategy is employed, combining NNS, NCE, and a risk assessor (RA) to create a bidirectional search path planning method, known as 3D NR-RRT. Starting from the starting point and the goal, the method gradually advances toward the intermediate region. This method is used to achieve collision avoidance and find a risk-bounded path in an uncertain environment. This algorithm uses NNS and NCE to bidirectionally search for a coarse global path. NNS expands the search tree with information, while NCE helps find a low-cost tree.

[0077] Step 4.1: RA is used to verify the path The collision risk of a given configuration node Location , if all risk profiles If the inequality in does not hold, the nodes in the risk area are deleted. Then, the collision risk of each edge between two adjacent nodes is verified by the SOS condition. There are three steps to verify the SOS condition using the spotless toolbox:

[0078] Step 1: Export linear trajectory , given the positions of two nodes as the endpoints of the line segment;

[0079] Step 2: List the time Semi-algebraic constraints , and created multiple multiplier variables ,in is a free variable, Polynomial coefficients for constructing and verifying path security;

[0080] Step 3: Mosek solver is used to minimize , which helps determine whether the state of the expression is primal feasible and dual feasible. If it does not exist, the line segment does not have the SOS form and the probability of a collision along it is greater than .

[0081] Step 4.2: Use the 3D NR-RRT algorithm with the input being the starting and target configuration pairs { , }, risk tolerance , and risk profile and its corresponding point cloud First, use PCnet to Encoded to In an environment with Initialized to empty. Create a forward node set containing the start state and the set of backward nodes with the target state . and The update process is the same. For each iteration, a batch of samples is generated by NNS. is the local starting point, NCE prediction Choose the one with the lowest cost. and add it to If and The connection of the top node in the equation satisfies the SOS condition, and it is considered that the node containing and Rough global path of the sample If not, an empty path is returned.

[0082] In getting Afterwards, LSC was used to remove The redundant nodes in the The non-adjacent nodes of the two nodes are used to form a shortcut segment. If the shortcut segment is within the safe area, the nodes between the two ends are removed. The remaining nodes may become part of the final solution. Then, the RA method is used to evaluate the path. And get the success mark , risk edge index , and the global path of the security node .

[0083] If the path succeeds and the risk is bounded, return the path Otherwise, use Replan to replan the risk segment and use LSC and RA methods to update the path. If a feasible path is found, return the updated path. .

Claims

1. A large-scale bulk material equipment path planning method based on risk boundary, characterized in that: The following steps are involved: Step 1: In an uncertain environment, first use Gazebo to build a bulk material equipment environment, then define the path planning problem, including the safe space, the uncertain dangerous environment, the safe initial state, the target state, and the acceptable risk level. Specific implementation steps are as follows: set up d = {2,3} represents the working space for large bulk material equipment operation; The danger zone is described as a polynomial in coordinate positions p∈ε: e obsi (oh i )={p∈ε:P i (p,ω i )≥0},i=1,...,n0, Among them, ε obsi (ω i ), i=1,...,n0 represents the static risk area in the workspace, is a given polynomial describing a risk region with uncertain size, location or geometry; set up It is the configuration space of bulk material equipment, where There are obstacles and risk is bounded x safe =x\x obs Configuration space, let x init and x goal As secure startup configuration and target configuration respectively; The goal of bulk equipment is to find the following risk-bounded path: π(t):[0,T]→x safe ,π(0)=x init ,π(T)=x goal , Where π(t)=[x init ,...,x goal ] means contains x safe Configuration and connection x init and x goal risky paths; The position of the path ψ(t)=[p init ,...,p goal ]:[0,T]→ε safe Given π(t)→x safe , the risk-bounded optimal path planning problem is defined as: Constraint π(0)=x0=x init ,π(T)=x n =x goal , Here, the cost function is the path length π(t) measured by Euclidean distance, and ∑ represents the set of all feasible paths; The probability constraint expressing the collision risk is: Where ψ(t) represents the position sequence on the path; Step 2: Create a risk contour map to describe the probabilistic information of the uncertain hazardous environment, divide the environment into safe areas, dangerous areas, and risk areas, use a neural network to encode the point cloud map, and convert the point cloud map into a potential embedding representation; specifically, follow the following steps to implement it: For each obstacle, an associated risk profile is defined, which represents the safe area around the obstacle at that risk level. Risk Profile Defined as the following set of states: The deterministic constraints on the safe state set are approximately: Where E[P(x,ω)] represents the mean value of the obstacle at position P, E[P 2 (x,ω)] represents the variance of the obstacle at position P; gather is the original risk profile An inner approximation based on rational polynomials, in the risk-bounded state set With a deterministic constraint on the probability of collision, the resulting path π consisting of these states is guaranteed to have no more than Δ collisions, and the risk profile is as follows: x(0)=x init ,x(T)∈X target , X target is the target state, Step 3: Build and train a neural network. NNS is trained to extract latent embedding features from point clouds and predict the state distribution in the latent subspace. NCE is used to build a neural network to evaluate the cost from the current state to the target state and select the best state. NNS includes a barrier encoder network PCnet and an inference network Inet. PCnet has four key building blocks for extracting the latent embedding Z, including a transformation network T-Net, CNN, a maximum pooling layer, and a multi-layer perceptron MLP. T-Net is a mini network that performs certain convolution and full connection operations on the original sample data. NCE consists of the same encoder and an encoder for cost prediction. NCE iteratively searches for the node with the lowest cost and selects the key state. Step 4: Adopt a bidirectional search strategy, combining NNS, NCE, and RA bidirectional search path planning method, namely 3D NR-RRT, starting from the starting point and the target and gradually moving towards the middle area, to achieve collision avoidance and find a risk-bounded path in an uncertain environment. The specific implementation is as follows: RA is used to verify the collision risk of path π, given the configuration node π i The SOS condition is used to verify the collision risk of each edge between two critical points at position p. There are three steps to verify the SOS condition using the spotless toolbox: Step 1: Substitution - Given the positions of two nodes as the endpoints of the line segment, substitute the trajectory into For each term of , we get the polynomial P with uncertain time t and derive the linear trajectory: {(x,y,z),x=a1t+b1,y=a2t+b2,z=a3t+b3}; Step 2: List the semi-algebraic constraint h = t * (1-t) about time t and create multiple multiplier variables: S=v1+v2t+v3t 2 +v4t 3 +v5t 4 +v6t 5 , where v i is a free variable, h is the polynomial coefficient for constructing and verifying the safety of the path; Step 3: Use the Mosek solver to minimize p-Sh; The 3D NR-RRT algorithm is used, and the input is the starting and target configuration pair {x init , x goal }, risk tolerance Δ, and risk profile and all internal approximations of its corresponding point cloud PM, First, use PCnet to encode PM into the environment encoded as Z, initialize the path solution π to empty, and create a forward node set V containing the start state. f and the backward node set V with the target state b ; V f and V b The updating process is the same; for each iteration, a batch of samples B is generated by NNS sample is the local starting point, NCE predicts B sample The cost of; choose the x with the lowest cost t1 and add it to V f In; if V f and V b If the connection of the top node in satisfies the SOS condition, it is considered that a node containing V is found. f and V b The coarse global path π of the sample; if not, an empty path is returned. After obtaining π, LSC is used to remove redundant nodes in V. It selects key nodes by trying to connect non-adjacent nodes of π. If the shortcut line is within the safe area, the nodes originally located between the two ends are removed, and the remaining nodes become part of the final solution. The RA method is used to evaluate the path π and obtain the success flag succ and the risk edge index F. idx , and the global path π of the security node, If the path is successful and the risk is bounded, return the path π. Otherwise, use Replan to replan the risk segment and update the path using LSC and RA methods. If a feasible path is found, return the updated path π.

Images