Energy-saving overturning control system for automatic obstacle avoidance of crane jib

Abstract

The invention belongs to the field of crane jib control, and particularly discloses an energy-saving turnover control system for automatic obstacle avoidance of a crane jib, which comprises a data acquisition module, an obstacle recognition module, an obstacle avoidance path design module and a turnover control module. According to the invention, through the Markov random field model, the gradient potential energy, the depth curvature potential energy and the depth variance potential energy are fused, the obstacle is segmented, and the features of the obstacle can be described more comprehensively, so that the segmentation accuracy is improved; a path planning method based on strategy gradient enhancement is used for designing a suspension arm path, the action and state of the suspension arm are continuously updated, explicit definition of an environment rule is not needed, the method is suitable for a dynamic and uncertain environment, long-term rewards are considered, and finding of a globally optimal or nearly optimal path is facilitated.

Description

Technical Field

[0001] The present invention relates to the field of crane boom control, and specifically refers to an energy-saving flipping control system for automatic obstacle avoidance of a crane boom. Background Art

[0002] Crane boom obstacle avoidance control can automatically detect the environment around the boom and adjust the angle and position of the boom according to the detection results to avoid collisions with obstacles, thereby improving the safety, efficiency, and reliability of the crane. When identifying obstacles around the crane boom, traditional methods usually require the use of multiple sensors and complex algorithms, which increases the complexity of the system and thus the difficulty of development and maintenance; when planning the boom path, traditional path planning methods are static and not applicable to dynamically changing environments. Summary of the Invention

[0003] In view of the above situation, to overcome the defects of the prior art, the present invention provides an energy-saving flipping control system for automatic obstacle avoidance of a crane boom. For the technical problem that when identifying obstacles around the crane boom, traditional methods usually require the use of multiple sensors and complex algorithms, which increases the complexity of the system and thus the difficulty of development and maintenance, the present invention uses a Markov random field model to fuse gradient potential energy, depth curvature potential energy, and depth variance potential energy to segment obstacles, which can more comprehensively describe the characteristics of obstacles and thus improve the accuracy of segmentation; for the technical problem that when planning the boom path, traditional path planning methods are static and not applicable to dynamically changing environments, the present invention uses a path planning method based on policy gradient reinforcement to design the boom path, continuously updating the boom actions and states, without explicitly defining environmental rules, applicable to dynamic and uncertain environments, and considering long-term rewards, which helps to find the global optimal or near-optimal path.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides an energy-saving flipping control system for automatic obstacle avoidance of a crane boom, and the energy-saving flipping control system for automatic obstacle avoidance of a crane boom includes a data acquisition module, an obstacle recognition module, an obstacle avoidance path design module, and a flipping control module; The data acquisition module installs a high-precision detachable distance sensor and an image sensor on the crane boom to obtain surrounding environment information in real time, and the environment information includes image information and angle information; The obstacle recognition module identifies obstacles in the environment information through a Markov random field model. If there are obstacles, the obstacle avoidance path design module plans the path of the crane boom, otherwise, the crane boom operates normally; The obstacle avoidance path design module constructs a boom path planning model and uses the policy gradient reinforcement method for training to output a low-energy-consuming boom path; The flipping control module designs the boom flipping control logic according to the boom path through the PID control algorithm, and uses the flipping device to control the output power and rotation speed of the motor for flipping. Specifically, the boom path is discretized into a time-angle sequence, where the angle is the target angle of the boom. The actual angle of the boom is collected every 10 ms using an angle encoder, and the difference between the actual angle and the target angle of the boom is calculated and recorded as the instantaneous error. The PID algorithm is used to generate a basic control quantity based on the instantaneous error, and the gravity compensation quantity is calculated according to the weight of the boom and the actual angle of the boom. The gravity compensation quantity is superimposed on the basic control quantity to obtain the total control quantity, and the total control quantity is converted into a PWM duty cycle. The flipping device controls the motor according to the PWM duty cycle.

[0005] Further, the obstacle recognition module recognizes obstacles in the environmental information through a visual recognition method. The visual recognition method specifically includes the following steps: Step A1: Collect an obstacle image set, establish and initialize a Markov random field model, input the obstacle image set into the Markov random field model for training, regard each pixel in the obstacle image as a node of the Markov random field model, perform unary and binary associations on each node, and define the cost function of the Markov random field model. The formula used is as follows: ; In the formula, is the cost function of the Markov random field model, is the set of random variables, and are the states of the pixels in the obstacle image, is the set of nodes of the Markov random field model, is the total number of pixels in the obstacle image, is the unary term, is the binary term; Step A2: Calculate the unary term of each pixel. The unary term is the product of the gradient potential energy, depth curvature potential energy, and depth variance potential energy of the obstacle image. The formula used is as follows: ; In the formula, is the unary term of the pixel, is the gradient potential energy, is the pixel coordinate of the obstacle image, and are the partial derivatives of the obstacle image in the horizontal and vertical directions respectively, is the depth curvature potential energy, , and are three features of the obstacle image, namely color histogram, gray-level co-occurrence matrix, and local binary, is the depth variance potential energy, is the depth image of the obstacle image, is at the disparity value, is the size of the obstacle image; Step A3: Calculate the binary term of each pixel, and the formula used is as follows: ; In the formula, is the binary term of the pixel, is the regularization parameter, is the Kronecker function; Step A4: Use a semantic-based image segmentation method to identify the obstacles in the obstacle image according to the unary term and the binary term, segment the obstacles in the obstacle image, identify the obstacle coordinates, and calculate the threat value of the obstacles. The formula used is as follows: ; In the formula, is the threat value of the obstacle, , and are the spatial coordinates of the obstacle, is the width of the obstacle; Step A5: When the threat value of the obstacle is greater than 0.3, it is regarded as an existing obstacle; otherwise, it is regarded as no obstacle.

[0006] Furthermore, the obstacle avoidance path design module uses a path planning method based on policy gradient reinforcement to design the boom path. The path planning method based on policy gradient reinforcement specifically includes the following steps: Step B1: Collect the crane boom path set, establish and initialize a boom path planning model. The boom path planning model is composed of a neural network, including a Q network, a deterministic policy network, a target Q network, and a target policy network. Input the crane boom path set into the boom path planning model for training. Define a trajectory tuple. The parameters of the trajectory tuple include the state, action, reward, and the state of the next time step at the current time step. Update the neural network parameters using the replay buffer R, and randomly extract N trajectory tuples from the replay buffer; Step B2: Train the Q network using the Bellman equation and use the Q network to learn the boom path planning. Define the objective function of the Q network, and the formula used is as follows: ; In the formula, is the objective function of the Q-network, is the state, is the reward, is the time step, is the discount factor in (0, 1), is the target Q-network, is the target policy network, is the learning parameter of the target policy network, is the learning parameter of the target Q-network. The Q-network is trained by minimizing the mean square loss between the main Q-function and the objective function, and the formula used is as follows: ; In the formula, is the loss function of the Q-network, is the number of randomly sampled trajectory tuples, is the action, is the learning parameter of the Q-network; Step B3: When performing the boom path planning, update the boom path using the policy gradient, and the formula used is as follows: ; In the formula, is the policy gradient, is the Q-network, is the policy network, is the learning parameter of the policy network; Step B4: Preset the update time step, and the target Q-network performs a mean update every update time step; Step B5: Determine the final boom path, and the formula used is as follows: ; In the formula, is the action at i.e., the boom path, is the state at is the learning parameter of the policy network at is the spatial noise at

[0007] The beneficial effects achieved by the present invention using the above solution are as follows: (1)In view of the technical problem that when identifying obstacles around the crane boom, traditional methods usually require the use of multiple sensors and complex algorithms, which increases the complexity of the system and thus the difficulty of development and maintenance, the present invention uses a Markov random field model to fuse gradient potential energy, depth curvature potential energy, and depth variance potential energy to segment obstacles, which can more comprehensively describe the characteristics of obstacles and thus improve the accuracy of segmentation; (2)In view of the technical problem that when planning the boom path, traditional path planning methods are static and not applicable to dynamically changing environments, the present invention designs the boom path using a path planning method based on policy gradient reinforcement, continuously updating the boom actions and states, without explicitly defining environmental rules, applicable to dynamic and uncertain environments, and considering long-term rewards, which helps to find the global optimal or near-optimal path. Description of the Drawings

[0008] Figure 1 It is a module connection diagram of an energy-saving flipping control system for automatic obstacle avoidance of a crane boom provided by the present invention.

[0009] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. Detailed Embodiments

[0010] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0011] Embodiment 1: Refer to Figure 1 , this embodiment provides an energy-saving flipping control system for automatic obstacle avoidance of a crane boom. The energy-saving flipping control system for automatic obstacle avoidance of a crane boom includes a data acquisition module, an obstacle recognition module, an obstacle avoidance path design module, and a flipping control module; The data acquisition module installs high-precision detachable distance sensors and image sensors on the crane boom to obtain surrounding environment information in real time. The environment information includes image information and angle information; The obstacle recognition module identifies obstacles in the environment information through a Markov random field model. If there are obstacles, the obstacle avoidance path design module plans the crane boom path. Otherwise, the crane boom operates normally; The obstacle avoidance path design module constructs a boom path planning model and uses the policy gradient reinforcement method for training to output a low-energy consumption boom path; The flipping control module designs the boom flipping control logic according to the boom path through the PID control algorithm, and uses the flipping device to control the output power and speed of the motor for flipping. Specifically, the boom path is discretized into a time-angle sequence, where the angle is the target angle of the boom. The actual angle of the boom is collected every 10 ms using an angle encoder, and the difference between the actual angle and the target angle of the boom is calculated and recorded as the instantaneous error. The PID algorithm is used to generate a basic control quantity based on the instantaneous error, and the gravity compensation quantity is calculated according to the weight of the boom and the actual angle of the boom. The gravity compensation quantity is superimposed on the basic control quantity to obtain the total control quantity, and the total control quantity is converted into a PWM duty cycle. The flipping device controls the motor according to the PWM duty cycle.

[0012] Example 2: Refer to Figure 1 , this example is based on the above example. When hoisting steel structure components at a construction site, the construction site environment is complex. The crane boom needs to detect obstacles in real time during operation, plan an obstacle avoidance path, and control the boom to complete the hoisting task safely and efficiently; The image sensor installed on the boom is a high-precision binocular sensor that captures the RGB-D image of the obstacle. The resolution of the RGB-D image is 1920×1080, which contains the color and depth information of the obstacle. The distance sensor installed on the boom is a lidar that measures the horizontal and vertical distances between the boom and the obstacle in real time; Semantic segmentation of the image is performed through the Markov random field model. When calculating the unary term of each pixel, the gradient potential energy of each pixel is obtained by calculating the horizontal and vertical gradients through the Sobel operator. The depth curvature potential energy is obtained by fusing the color histogram, gray-level co-occurrence matrix, and local binary features. The depth variance potential energy is obtained by calculating the dispersion degree of the local depth values. When calculating the binary term of each pixel, the relevance of adjacent pixels is combined, and the Kronecker function is used to determine whether the pixel belongs to the same obstacle area, and the regularization parameter is set to 0.8. According to the semantic segmentation result, the threat value of the obstacle is calculated. If it has a high threat, an obstacle avoidance path planning is immediately carried out; The policy gradient reinforcement method is used to generate a low-energy consumption path. In the training stage, the crane boom path set is used, and the crane boom path set contains 1000 paths. The discount factor is set to 0.95, and the capacity of the replay buffer is 10,000 trajectory tuples. Actions are output through the policy network to explore the optimal path. The Q network is used to evaluate the long-term reward of the action, and the Q value is updated using the Bellman equation. Finally, a final path bypassing the obstacle is planned.

[0013] Example 3: Refer toFigure 1 , this embodiment is based on the above embodiment. The obstacle recognition module recognizes obstacles in the environmental information through a visual recognition method. The visual recognition method specifically includes the following steps: Step A1: Collect an obstacle image set, establish and initialize a Markov random field model, input the obstacle image set into the Markov random field model for training, regard each pixel in the obstacle image as a node of the Markov random field model, perform unary and binary associations on each node, and define the cost function of the Markov random field model. The formula used is as follows: ; In the formula, is the cost function of the Markov random field model, is a set of random variables, and are the states of the pixels in the obstacle image, is the set of nodes of the Markov random field model, is the total number of pixels in the obstacle image, is the unary term, is the binary term; Step A2: Calculate the unary term of each pixel. The unary term is the product of the gradient potential energy, depth curvature potential energy, and depth variance potential energy of the obstacle image. The formula used is as follows: ; In the formula, is the unary term of the pixel, is the gradient potential energy, is the pixel coordinate of the obstacle image, and are the partial derivatives of the obstacle image in the horizontal and vertical directions respectively, is the depth curvature potential energy, , and are three features of the obstacle image, namely color histogram, gray-level co-occurrence matrix, and local binary, is the depth variance potential energy, is the depth image of the obstacle image, is at the disparity value, is the size of the obstacle image; Step A3: Calculate the binary term of each pixel. The formula used is as follows: ; In the formula, is the binary term of the pixel, is the regularization parameter, is the Kronecker function; Step A4: Use a semantic-based image segmentation method to identify obstacles in the obstacle image according to unary terms and binary terms, segment the obstacles in the obstacle image, identify the obstacle coordinates, and calculate the threat value of the obstacle. The formula used is as follows: ; In the formula, is the threat value of the obstacle, , and are the spatial coordinates of the obstacle, is the width of the obstacle; Step A5: When the threat value of the obstacle is greater than 0.3, it is considered that there is an obstacle; otherwise, it is considered that there is no obstacle.

[0014] Through the above operations, when identifying obstacles around the crane boom, traditional methods usually require the use of multiple sensors and complex algorithms, which will increase the complexity of the system and thus increase the difficulty of development and maintenance. The present invention uses a Markov random field model to fuse gradient potential energy, depth curvature potential energy, and depth variance potential energy to segment obstacles, which can more comprehensively describe the characteristics of obstacles and thus improve the accuracy of segmentation.

[0015] Example 4: Refer to Figure 1 , this example is based on the above example. The obstacle avoidance path design module uses a path planning method based on policy gradient reinforcement to design the boom path. The path planning method based on policy gradient reinforcement specifically includes the following steps: Step B1: Collect the crane boom path set, establish and initialize a boom path planning model. The boom path planning model is composed of a neural network, including a Q network, a deterministic policy network, a target Q network, and a target policy network. Input the crane boom path set into the boom path planning model for training. Define a trajectory tuple. The parameters of the trajectory tuple include the state, action, reward, and the state of the next time step at the current time step. Use the replay buffer R to update the neural network parameters, and randomly extract N trajectory tuples from the replay buffer; Step B2: Use the Bellman equation to train the Q network and use the Q network to learn the boom path planning. Define the objective function of the Q network. The formula used is as follows: ; In the formula, is the objective function of the Q network, is the state, is the reward, is the time step, is the discount factor of (0, 1), is the target Q-network, is the target policy network, are the learning parameters of the target policy network, are the learning parameters of the target Q-network. The Q-network is trained by minimizing the mean squared loss between the main Q-function and the target function. The formula used is as follows: ; In the formula, is the loss function of the Q-network, is the number of randomly sampled trajectory tuples, is the action, are the learning parameters of the Q-network; Step B3: When performing the boom path planning, update the boom path using policy gradient. The formula used is as follows: ; In the formula, is the policy gradient, is the Q-network, is the policy network, are the learning parameters of the policy network; Step B4: Preset the update time step. The target Q-network performs a mean update every update time step; Step B5: Determine the final boom path. The formula used is as follows: ; In the formula, is the action at i.e., the boom path, is the state at is the learning parameters of the policy network at is the spatial noise at

[0016] Through the above operations, for the technical problem that in the boom path planning, the traditional path planning method is static and not applicable to the dynamically changing environment, the present invention designs the boom path using the path planning method based on policy gradient reinforcement, continuously updates the boom actions and states, does not require explicit definition of environmental rules, is applicable to dynamic and uncertain environments, and considers long-term rewards, which helps to find the global optimal or near-optimal path.

[0017] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0018] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

[0019] The above description of the present invention and its embodiments is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of the present invention, design similar structural forms and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. An energy-saving flipping control system for automatic obstacle avoidance of a crane boom, characterized in that It includes a data acquisition module, an obstacle recognition module, an obstacle avoidance path design module, and a flipping control module; The data acquisition module installs a high-precision detachable distance sensor and an image sensor on the crane boom to obtain real-time surrounding environment information; The obstacle recognition module recognizes obstacles in the environmental information through a Markov random field model. If there are obstacles, the obstacle avoidance path design module plans the path of the crane boom. Otherwise, the crane boom works normally; The obstacle avoidance path design module constructs a boom path planning model and uses the policy gradient reinforcement method for training to output a low-energy-consuming boom path; The flipping control module designs the boom flipping control logic according to the boom path through the PID control algorithm, and uses the flipping device to control the output power and speed of the motor for flipping. Specifically, the boom path is discretized into a time-angle sequence, where the angle is the target angle of the boom. The actual angle of the boom is collected using an angle encoder, and the difference between the actual angle and the target angle of the boom is calculated and recorded as the instantaneous error. The PID algorithm is used to generate a basic control quantity based on the instantaneous error. The gravity compensation quantity is calculated according to the weight of the boom and the actual angle of the boom, and the gravity compensation quantity is superimposed on the basic control quantity to obtain the total control quantity. The total control quantity is converted into a PWM duty cycle, and the flipping device controls the motor according to the PWM duty cycle.

2. The energy-saving flipping control system for automatic obstacle avoidance of a crane boom according to claim 1, characterized in that, The obstacle recognition module recognizes obstacles in the environmental information through a visual recognition method. The visual recognition method specifically includes the following steps: Step A1: Collect an obstacle image set, establish and initialize a Markov random field model, input the obstacle image set into the Markov random field model for training, and define the cost function of the Markov random field model. The formula used is as follows: ； In the formula, is the cost function of the Markov random field model, is a set of random variables, and are the states of the pixels in the obstacle image, is the set of nodes of the Markov random field model, is the total number of pixels in the obstacle image, is the unary term, is the binary term; Step A2: Calculate the unary term of each pixel. The unary term is the product of the gradient potential energy, depth curvature potential energy, and depth variance potential energy of the obstacle image; Step A3: Calculate the binary term of each pixel through the Kronecker function; Step A4: Use a semantic-based image segmentation method to segment the obstacles in the obstacle image, identify the obstacle coordinates, and calculate the threat value of the obstacles. The formula used is as follows: ； Wherein, is the threat value of the obstacle, , and are the spatial coordinates of the obstacle, is the width of the obstacle; Step A5: When the threat value of the obstacle is greater than 0.3, it is considered that there are obstacles. Otherwise, it is considered that there are no obstacles.

3. The energy-saving flipping control system for automatic obstacle avoidance of a crane boom according to claim 2, characterized in that, The obstacle avoidance path design module designs the boom path using a path planning method based on policy gradient reinforcement. The path planning method based on policy gradient reinforcement specifically includes the following steps: Step B1: Collect a crane boom path set, establish and initialize a boom path planning model. The boom path planning model is composed of a neural network, including a Q network, a deterministic policy network, a target Q network, and a target policy network. Input the crane boom path set into the boom path planning model for training. Define a trajectory tuple. The parameters of the trajectory tuple include the state, action, reward, and next time step state at the current time step. Update the neural network parameters using the replay buffer R; Step B2: Train the Q-network using the Bellman equation, use the Q-network to learn the boom path planning, define the objective function of the Q-network, and the formula used is as follows: ； where is the objective function of the Q-network, is the state, is the reward, is the time step, is the discount factor in (0, 1), is the target Q-network, is the target policy network, are the learning parameters of the target policy network, are the learning parameters of the target Q-network. The Q-network is trained by minimizing the mean squared loss between the main Q-function and the objective function; Step B3: Update the boom path using the policy gradient, and the formula used is as follows: ； Wherein, is the policy gradient, is the Q-network, is the policy network, are the learning parameters of the policy network, are the learning parameters of the Q-network, is the number of randomly sampled trajectory tuples; Step B4: Preset the update time step, and the target Q-network performs a mean update every update time step; Step B5: Determine the final boom path, and the formula used is as follows: ； wherein, is the action at i.e., the boom path, is the state at the learning parameter of the policy network at is the spatial noise at

Images