Insulator wiping curved arm sequence control method based on digital twinning

Abstract

This invention belongs to the field of robot path planning and intelligent control technology, specifically relating to a sequential control method for an insulator wiping arm based on digital twins. The method includes: Step 1, constructing a digital twin platform with constrained space; Step 2, training a deep reinforcement learning policy network to generate motion commands adapted to the sequential execution mode; Step 3, generating a discrete-time joint angle sequence based on the deep reinforcement learning policy network; Step 4, performing pose safety verification and collision detection on the discrete-time joint angle sequence based on the digital twin platform; and Step 5, performing sequential execution of the virtual-real closed loop based on the verified discrete-time joint angle sequence. The advantages of this invention include: overcoming modeling bottlenecks with high-fidelity dynamic mapping; enhancing obstacle avoidance capabilities and path search robustness in complex environments; constructing a hard fit for physical constraints to ensure inherent safety during operation; and achieving high-precision correction and dynamic smoothing compensation in the virtual-real closed loop.

Description

Technical Field

[0001] This invention relates to a sequential control method for insulator wiping arm based on digital twins, belonging to the field of robot path planning and intelligent control technology. Background Technology

[0002] With the development of electrified railways towards intelligent operation and maintenance, using robots to replace manual labor for overhead contact line maintenance has become an inevitable trend. The overhead contact line environment is characterized by limited space and high precision requirements. To address the specific needs of rail transit overhead contact line operations, which involve limited space and complex obstacles, a customized 7-DOF non-standard curved arm insulator for wiping the overhead contact line cantilever arm has been developed.

[0003] The boom arm has a large working radius of 4100mm in the horizontal plane and a working height of 8500mm in the vertical direction. Its hardware architecture employs an electro-hydraulic hybrid drive system, powered independently by a lifting pump station and an upper pump station. Due to the response differences of the heterogeneous power systems and the constraints of the control architecture, the boom arm must follow a sequential control mode where joints advance in a preset time sequence during actual operation, making synchronous interpolation of multiple joints impossible. However, as this boom arm is a non-standard customized device, its complex linkage structure, the intermediate attitude uncertainties arising from the sequential execution logic, and the requirements for flexible high-altitude operations pose significant challenges to traditional modeling and motion control.

[0004] Digital twin and path planning technologies have developed rapidly, but the following shortcomings still exist in the control of articulated booms for overhead contact line operation and maintenance: Difficulty in solving kinematics: Because the curved arm is a customized 7-DOF structure, and it is difficult to obtain extremely accurate DH (Denavit-Hartenberg) parameters during manufacturing and assembly, traditional analytical or numerical methods have a huge computational load and cannot guarantee accuracy when solving inverse kinematics, which cannot meet the requirements of real-time operation.

[0005] Collision risk in sequential execution mode: Constrained by the control architecture and power system response characteristics, the articulated boom must follow the logic of joints moving sequentially in a preset order during execution. Traditional path planning algorithms are often based on the ideal assumption of multi-axis synchronous execution, which leads to a large number of intermediate postures that are not considered in the planning stage during actual execution. Even if the planned path shows no collision in the standard simulation environment, the articulated boom is very likely to collide unexpectedly with the overhead contact line or target equipment in the sequential action sequence of the physical entity. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the prior art by providing a digital twin-based method for controlling the sequential wiping of insulator arms.

[0007] The technical solution provided by this invention to solve the above-mentioned technical problems is: a method for sequential control of insulator wiping arm based on digital twin, comprising the following steps: Step 1: Construct a digital twin platform with constrained space; Step 2: Train a deep reinforcement learning policy network to generate motion instructions that adapt to the sequential execution pattern; Step 3: Generate discrete temporal joint angle sequences based on a deep reinforcement learning policy network; Step 4: Perform pose safety verification and collision detection on discrete temporal joint angle sequences based on the digital twin platform; Step 5: Perform sequential execution of the virtual and real closed loop based on the verified discrete time-series joint angle sequence.

[0008] A further technical solution is that the specific process of step 1 includes: Step 11: Construct a digital articulated boom model in the digital twin platform based on the actual structural parameters of the boom; Step 12: Based on the structural parameters and working environment of the work area, construct a working environment model including obstacles and safety constraints in the digital twin platform; Step 13: Establish a communication mechanism based on the robot operating system between the digital twin platform and the physical articulated arm control system.

[0009] A further technical solution is that the working environment model includes a digital scene comprising maintenance vehicles, railway tracks, overhead contact line columns, overhead contact line cantilever arms, cantilever arm support pipes, positioners, contact wires, catenary wires, droppers, and high-voltage conductors.

[0010] A further technical solution is that the operating environment model configures high-precision collision body attributes for each component, and according to the electrified railway safety operation procedures, the boom and wiping tool must maintain a minimum safe distance from the contact line and catenary. At the same time, a safe distance constraint space is constructed by extending outward from the contact line as a reference. The constraint space is transformed into a collision penalty field in the physics simulation engine, so that the sweep volume of the boom body and the end wiping tool in the virtual environment is monitored in real time, thereby providing accurate environmental geometric boundaries and safe clearance constraints for subsequent motion planning.

[0011] A further technical solution is that the input to the deep reinforcement learning policy network includes a composite state observation vector. ;

[0012]

[0013]

[0014]

[0015]

[0016]

[0017] In the formula: The state vector of the curved arm body; It is a set of generalized joint coordinates; The rotation angle; and These are the position coordinates and attitude quaternions of the end effector in three-dimensional space, respectively; Guide the feature vector for the target; The coordinates of the center of the target insulator. The distance between the end effector and the target point is the Euclidean distance. The unit direction vector that guides the movement of the end effector; Features for sensing environmental obstacles; Here are the coordinates of the obstacle's position. For the first bent arm Link and the first The shortest distance between obstacles The minimum safe clearance between the end effector and the obstacle; This is a vector indicating the active state of a joint; This is the index number of the active joint currently determined based on timing logic.

[0018] A further technical solution is that the deep reinforcement learning policy network constructs action mask logic at the policy inference layer, the mathematical expression of which is:

[0019] In the formula: This is the index number of the active joint currently determined based on timing logic; This is the final execution instruction after masking correction; The original action vector output by the policy network; This is the action mask vector.

[0020] A further technical solution is that the reward function of the deep reinforcement learning policy network... Rewards from mission objectives Safety boundary penalties and dynamic smoothing penalty It consists of three parts, used to guide the SAC strategy model to achieve safe and accurate path search during parallel simulation;

[0021]

[0022] In the formula: This represents the negative exponential distance between the end effector and the target point; The Euclidean distance between the current position of the end effector and the target wiping point; The safety boundary penalty is designed with a distance for the overhead contact line cantilever and surrounding obstacles; This is the shortest distance between the edge of the swept volume and the contact line; This is the preset minimum safety gap threshold; This is a penalty term for a positive constant. The dynamic smoothing penalty is designed to address the sensitivity of electro-hydraulic hybrid drive systems to abrupt changes in motion. This represents the effective action quantity at the current moment after being corrected by the masking operator.

[0023] A further technical solution is that the specific process of step 3 includes: Step 31: Solidify the parameters of the converged policy network during training to build an offline inference engine; Step 32: Obtain the three-dimensional coordinates of the insulator to be wiped in the robot coordinate system, and use them as the target guidance feature vector. The input reference is used; simultaneously, the physical poses of each joint of the articulated arm are collected as the initial state to initialize the composite state observation vector. The state vector of the curved arm body ; Step 33: The path planner inputs the current composite state observation vector into the fixed strategy model, uses the joint active state indicator vector in it to identify the motion stage, and iteratively calculates the target pose increment of the current active joint by combining the action masking mechanism. Based on this, it updates and generates a complete 7-dimensional joint space pose node. Then, it realizes the triggering of the next stage action and the construction of discrete sequence by automatically migrating the joint active state indicator bit. Step 34, through The process involves sequential, iterative reasoning rounds until all seven joints have completed their target searches, ultimately constructing a framework containing... A complete linked list of discrete job paths for an ordered sequence of nodes; Step 35: Perform a smoothness check and normalization inverse calculation on the generated angle sequence, and restore the output normalized values ​​to radians and displacement units that can be recognized by the physical layer.

[0024] A further technical solution is that the specific process of step 4 includes: Step 41, generate the included The complete path list of each pose node is imported into the digital twin simulation platform, and the high-fidelity virtual articulated arm model is driven sequentially according to the ordered nodes in the list to restore the entire process trajectory of the physical entity from the initial pose to the target work position in the digital space. Step 42: For the single-axis progressive motion in each round of the curved arm, calculate the sweep volume formed by the motion envelope between adjacent pose nodes; Step 43: Monitor the geometric feature relationship between the swept volume and the obstacle constraint space constructed in the first step in real time, and verify the path and check the safety distance; Step 44: Perform segmented security assessment on the complete path list; if the intermediate sweep volume is detected to intrude into the constraint space described in Step 1, immediately execute the instruction interception logic, block the issuance permission of the path segment, and feed back the index and coordinate data of the interference node to the policy network in Step 2 in real time for path replanning; if the entire path passes the geometric verification, mark the path list as a safe and executable state, and use it as the final compliance benchmark for the actions of the physical execution mechanism.

[0025] A further technical solution is that the specific process of step 5 includes: Step 51: Using the communication plugin provided by the digital twin simulation platform, the verified discrete time-series angle sequence is encapsulated into control messages in sequence; using a distributed publish-subscribe mechanism, the pose nodes are distributed to the physical articulated boom control system in stages; according to the instructions, the physical end drives the motor and hydraulic pump station to complete the joint actions in physical order; if a sudden collision risk or inconsistent instructions occur at the physical layer, the digital twin system immediately triggers the forced interrupt logic, cuts off the physical drive source and locks all joints; Step 52: Using encoders and pose sensing sensors deployed on the physical articulated arm, the actual generalized coordinates and end-effector poses of each joint are collected in real time; the data is fed back to the digital twin platform in real time through the communication link to drive the virtual model to reconstruct the pose. Step 53: Calculate the error vector between the preset pose node of the digital twin and the real-time pose feedback; If the deviation exceeds the preset threshold, the system will use a closed-loop control algorithm or feed the residual back to the strategy network; calculate the corrected joint target value in real time, and add the compensated increment to the control signal of the currently active joint to achieve real-time correction of the physical end action; Step 54: After each round of sequential execution, the system determines whether the end effector has achieved the preset target through pose feedback; If the task is completed, the system generates a work completion beacon and drives the articulated boom to retract to its initial position along a preset path, completing the entire virtual-real closed-loop sequential work process.

[0026] The present invention has the following beneficial effects: 1. High-fidelity dynamic mapping breaks through modeling bottlenecks; By constructing state observation vectors An end-to-end depth mapping between joint space and Cartesian space pose was established. This design allows the SAC algorithm to bypass the inverse kinematics calculations that rely on DH parameters in traditional robot control, effectively solving the problem of model inaccuracies caused by the electro-hydraulic actuation characteristics of the customized 7-DOF articulated arm. Combined with a digital twin simulation platform integrating the PhysX advanced physics engine, this invention can faithfully reproduce the kinematic response of the articulated arm in a physical environment, providing a reliable digital logic benchmark for subsequent precision operations.

[0027] 2. Enhance obstacle avoidance capabilities and path search robustness in complex environments; Based on the maximum entropy reinforcement learning framework, this invention significantly improves the robot's optimization ability in the unstructured environment of railway overhead contact lines. The SAC algorithm, by maximizing policy entropy, encourages the curved arm to explore diverse areas within narrow wrist-arm gaps, effectively avoiding the problem of traditional algorithms easily getting trapped in local optima. Simultaneously, a temporal logical vector identifying joint active states is introduced into the state space. This enables the model to accurately identify motion phases, eliminates state ambiguity in single-axis sequential execution, and ensures extremely high robustness of path generation under complex obstacle constraints.

[0028] 3. Construct physical constraint hard adaptation to ensure inherent safety during operation; The motion masking mechanism and sweep volume verification process, from the underlying logic to the pre-simulation stage, dually safeguard the electrical safety of railway operations. By forcibly intervening through the motion masking operator output, the algorithm-generated instructions strictly adhere to the single-axis progressive execution sequence of the joints at the physical level, eliminating the risk of accidental interference that may arise from multi-axis linkage. Furthermore, by utilizing a digital twin platform to calculate the sweep volume between adjacent pose nodes, a full collision scan of the joint motion transition process is achieved, ensuring that the articulated boom body and the contact line always maintain a safe clearance.

[0029] 4. Achieve high-precision correction and dynamic smoothing compensation in the virtual-real closed loop; The virtual-physical bidirectional mapping link effectively compensates for potential pose residuals that may occur in the electro-hydraulic drive system during actual operation. The closed-loop feedback system, comprised of the safety risk assessment module and the SAC strategy model, can intelligently calculate pose errors transmitted back in real-time from the physical end and use dynamic data provided by the PhysX engine to calculate the optimal compensation increment. Since the compensation action is also constrained by the maximum entropy objective function and only applies to the currently active joint, the entire correction process exhibits extremely high smoothness and stability, ensuring the insulator wiping accuracy while avoiding instantaneous pressure shocks to the physical drive mechanism. Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention; Figure 2 A diagram illustrating the sequential control method for insulator wiping arm based on digital twins; Figure 3 A schematic diagram of the operation of wiping insulators with a curved arm; Figure 4 This is a schematic diagram of the articulated arm structure. Detailed Implementation

[0031] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, the present invention provides a method for sequential control of insulator wiping arm sequence based on digital twins. The specific objectives of this method are as follows: 1. By leveraging the end-to-end learning capabilities of deep reinforcement learning, control strategies can be obtained through trial and error in a digital twin environment using reinforcement learning methods. Inverse kinematics can be solved without relying on high-precision DH parameters, effectively solving the control problem of the articulated arm.

[0033] 2. Construct a planning model with joint sequence control capability: Utilize the self-learning capability of deep reinforcement learning in the environment, embed sequential execution constraints into the training operator, so that the path generated by the model adapts to the sequential motion sequence of the joint axes, and eliminates the collision risk caused by unplanned intermediate poses from the bottom layer of the algorithm.

[0034] 3. Construct a feedforward safety verification mechanism: Utilize the uniqueness and predictability of the sequential execution trajectory to conduct collision pre-simulation in virtual space. Through the high-precision obstacle avoidance feedback provided by the digital twin, ensure that the articulated boom maintains an absolute safe distance from the contact line at every intermediate posture throughout the entire operation process.

[0035] 4. Real-time deviation compensation in a virtual-physical closed loop: Based on the two-way interactive characteristics of the digital twin system, the pose deviation of the physical entity during sequential execution is monitored in real time. Feedback information from the virtual model is used to dynamically correct and enable the sequential actions in the next stage, ensuring the stability and safety of the boom lift operation.

[0036] Specifically, the following steps are included: Step 1: Construct a digital twin platform with constrained space; Step 11: Construct a digital articulated arm model in the digital twin platform based on the actual structural parameters of the articulated arm. This model is used to describe the connection relationship, joint type, range of motion and driving characteristics between the links of the articulated arm, and to model the geometric shape, mass distribution, moment of inertia and self-locking stiffness characteristics introduced by the hydraulic system of the articulated arm. Meanwhile, the control methods of each joint in the digital articulated arm model are constrained so that they can only move in a preset order, thereby ensuring that the digital twin model can truly reflect the motion behavior and intermediate posture changes of the articulated arm in the sequential execution control mode. The structure of the articulated arm is as follows: Figure 4 As shown, it specifically includes 7 joints, which execute control logic in sequence, including the first rotary joint, the first translational joint, the second rotary joint, the second translational joint, the third rotary joint, the fourth rotary joint, and the fifth rotary joint. The end effector is an opening and closing brush.

[0037] The articulated boom is powered by an electro-hydraulic hybrid drive system. The first rotary joint is driven by a motor, the first translational joint is driven by a lifting pump station hydraulically, and the remaining joints are driven by an upper pump station hydraulically. This hardware requirement dictates that the articulated boom must execute in a predetermined sequence.

[0038] Step 12: Based on the structural parameters and working environment of the work area, construct a working environment model containing obstacles and safety constraints in the digital twin platform to provide basic environmental constraints for subsequent sequential control strategy training and path planning.

[0039] The operational environment model includes a digital scene encompassing the maintenance vehicle, railway track, overhead contact line posts, overhead contact line cantilever arms, cantilever arm support pipes, positioners, contact wires, catenary cables, droppers, and high-voltage conductors. High-precision collision attributes are configured for each component within the scene, and in accordance with electrified railway safety operating procedures, the cantilever arms and wiping tools must maintain a minimum safe distance from the contact wires and catenary cables. A safe distance constraint space is constructed by extending outward from the contact line as a reference.

[0040] The constraint space is transformed into a collision penalty field in the physics simulation engine, so that the sweep volume of the curved arm body and the end wiping tool in the virtual environment can be monitored in real time, thereby providing accurate environmental geometric boundaries and safety clearance constraints for subsequent motion planning. Step 13: Establish a communication mechanism based on the robot operating system between the digital twin platform and the physical articulated arm control system to realize the state interaction and control coordination between the virtual model and the physical entity during the sequential execution process; The communication mechanism sends the joint states, end-effector poses, execution feedback, and safety status of the physical articulated arm during execution to the digital twin environment via a publish-subscribe approach, thereby achieving synchronization and timing alignment between the virtual and physical states. Simultaneously, the discrete timing joint angle sequences generated by the digital twin environment are distributed to the physical articulated arm control system in stages through service calls or action communication, with stage completion flags and safety status serving as trigger conditions for subsequent action distribution.

[0041] Step 2: Train a deep reinforcement learning policy network to generate motion instructions that adapt to the sequential execution pattern; This embodiment uses the Isaac Lab reinforcement learning framework to construct a large-scale parallel simulation environment. Leveraging its native support for GPU tensor quantization computation, it enables parallel evolution of massive random pose samples within a virtual space. In the constructed railway operation scenario, the articulated boom simultaneously performs obstacle avoidance exploration under sequential control constraints across thousands of parallel simulation instances, thereby generating a policy network model with extremely high robustness and generalization ability. To address path planning problems in confined spaces, a Soft Actor-Critic (SAC) algorithm is introduced to construct the core of intelligent decision-making. Unlike traditional reinforcement learning that focuses on a single reward, this algorithm incorporates policy entropy into the optimization objective, enabling the curved arm to possess stronger exploration capabilities and robustness in narrow wrist-arm gaps.

[0042] It follows the maximum entropy reinforcement learning objective function Defined as:

[0043] In the formula: This refers to the comprehensive reward signal obtained after performing the bent-arm movement; The richness of the strategy distribution; This is a temperature regulation parameter used to balance the weights of work efficiency and motion exploration, ensuring that the model can find the optimal path to avoid obstacles.

[0044] The input to this deep reinforcement learning policy network is a composite state observation vector. ;

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] In the formula: The state vector of the curved arm body; For generalized joint coordinates, For a translational joint, the rotation angle is... This is a translational displacement; and These are the position coordinates and attitude quaternions of the end effector in three-dimensional space, respectively; Guide the feature vector for the target; The coordinates of the center of the target insulator. The distance between the end effector and the target point is the Euclidean distance. The unit direction vector that guides the movement of the end effector; Features for sensing environmental obstacles; Here are the coordinates of the obstacle's position. For the first bent arm Link and the first The shortest distance between obstacles The minimum safe clearance between the end effector and the obstacle; This is a vector indicating the active state of a joint; This is the active joint index number determined based on timing logic. If the current step specifies the first... Joint movement, The rest are 0, thus eliminating the state ambiguity of multi-joint asynchronous motion.

[0051] To address the physical constraints of the uniaxial progressive motion of the articulated arm hardware joint, an action mask logic is constructed at the policy reasoning layer, mathematically expressed as follows:

[0052] In the formula: This represents the original action vector output by the policy network, containing the expected changes at 7 joints. Action mask vector. Its dimensions are consistent with the number of joints. The active joint index number currently determined based on timing logic. This indicates that corresponding elements are multiplied. This indicates the final execution instruction after masking correction.

[0053] This operator ensures that at any given time, the pose increment of all joints except the active joints is 0, thus guaranteeing that the strategy satisfies the sequential constraint of joint execution.

[0054] The reward function of deep reinforcement learning policy networks Rewards from mission objectives Safety boundary penalties and dynamic smoothing penalty It consists of three parts, used to guide the SAC strategy model to achieve safe and accurate path search during parallel simulation;

[0055]

[0056] in This represents the negative exponential distance between the end effector and the target point, designed to incentivize the articulated boom end effector to approach the target insulator position. This is the Euclidean distance between the current position of the end effector and the target wiping point. As the distance reduction factor increases, this reward value grows exponentially, powerfully guiding the articulated boom to find the optimal trend to reach the work position within the complex cantilever structure.

[0057] The safety boundary penalty is designed for the distance of the overhead contact line cantilever and surrounding obstacles. This is the shortest distance between the edge of the swept volume and the contact line. Less than or equal to the preset minimum safety gap threshold Or, in the event of a hard collision involving the articulated boom body, the system assigns an extremely large negative penalty value. ( (This is a positive constant penalty term), and this hard constraint penalty forces the model to avoid any poses that may lead to flashover or collision during training.

[0058] When the distance is within a safe range, this invention introduces a logarithmic guided reward. , This represents the safety guidance weight coefficient. It is a positive value, and it increases with distance. As the value increases, the reward value increases smoothly, but the growth rate gradually slows down. This encourages the strategy to keep the curved arm as far away from the contact line as possible, while meeting the basic safety threshold, thereby leaving more safety redundancy during physical execution.

[0059] To address the sensitivity of electro-hydraulic hybrid drive systems to abrupt changes in motion, a smoothness constraint is introduced as a dynamic smoothing penalty. This represents the effective motion quantity at the current moment after correction by the masking operator. By limiting the fluctuation amplitude of joint commands between adjacent moments, the impact load on the hydraulic pump station and motor during sequential switching is reduced, thereby improving the operational stability and service life of the physical entity.

[0060] By adjusting the weight coefficients The size of the [device / device] can dynamically balance the priorities of efficiency and safety at different stages of the operation. For example, in the delicate work area near the contact line, by increasing [the size / device]... The value of makes the model exhibit extremely high obstacle avoidance sensitivity.

[0061] Step 3: Generate discrete temporal joint angle sequences based on a deep reinforcement learning policy network; Step 31: Solidify the parameters of the converged policy network during training to build an offline inference engine; When performing inference tasks, the strategy is changed from random distribution sampling in the training phase to deterministic mapping, that is, the mean of the Gaussian distribution is taken as the output action to ensure that the generated task path has high reproducibility and motion stability. Step 32: Obtain the three-dimensional coordinates of the insulator to be wiped in the robot coordinate system, and use them as the target guidance feature vector. The input reference is used; simultaneously, the physical poses of each joint of the articulated arm are collected as the initial state to initialize the composite state observation vector. The state vector of the curved arm body ; At this time, the joint active state indicator vector Initialize to [1, 0, 0, 0, 0, 0, 0], indicating that the first rotational joint is the active axis to be planned; Step 33: The path planner will use the current composite state observation vector The input is fed into the fixed strategy model, utilizing the joint active state identifier vector within it. The motion phase is identified, and the target pose increment of the currently active joint is iteratively calculated by combining the action masking mechanism. Based on this, a complete 7-dimensional joint spatial pose node is generated. Then, the triggering of the next stage of action and the construction of discrete sequences are realized by automatically migrating the joint active state flag bits. For complex obstacle avoidance paths, the path planner dynamically decomposes the task path into multiple sets of 7-DOF pose arrays through multi-round ordered cyclic reasoning, enabling the end effector to approximate the target position in segments. Within each execution cycle, the planner updates the current composite state observation vector... The input is fed into a fixed strategy operator, which uses a temporal logic component in the form of joint active state identifiers to accurately identify the current motion stage of the bent arm. Logic blocking ensures that the output of inactive joints is zero, and only the target increment of the currently active joint in the current round is calculated. Subsequently, the system updates the active axis coordinates based on the inferred increment while keeping other axes stationary, generating a complete 7D joint space pose node for the current stage. And automatically migrate after the joint completes pose update. The enable bit is set, and this node is used as the starting state to trigger the next stage of reasoning, until the sequential planning of all rounds is completed.

[0062] Step 34, through The process involves sequential, iterative reasoning rounds until all seven joints have completed their target searches, ultimately constructing a framework containing... A discrete sequence of complete job path linked list of ordered nodes ;

[0063] Each group Representing the A subset of attitude nodes formed after sequential execution of rounds. Each node All precisely define the first... Round The global static attitude after the axis motion is completed.

[0064] Step 35: Perform smoothness checks and normalization inverse calculations on the generated angle sequence, and restore the output normalized values ​​to radians and displacement units that the physical layer can recognize. Step 4: Perform pose safety verification and collision detection on discrete temporal joint angle sequences based on the digital twin platform; Step 41, generate the included The complete path list of each pose node is imported into the digital twin simulation platform, and the ordered nodes in the list are used to... The high-fidelity virtual articulated boom model is driven sequentially to recreate the entire trajectory of the physical entity from its initial pose to the target work position in the digital space. Step 42: For the single-axis progressive motion of the curved arm in each round, calculate the adjacent pose nodes. and The swept volume formed by the motion envelope between them; The swept volume is simplified to the geometric envelope generated by a single active joint motion, which significantly reduces the computational cost of interferometry verification.

[0065] Step 43: Monitor the geometric feature relationship between the swept volume and the obstacle constraint space constructed in the first step in real time, and verify the path and check the safety distance; The verification process includes: verifying whether there is spatial overlap between the articulated boom connecting rod and the contact wire support, cantilever, etc.; and dynamically monitoring the shortest distance between the edge of the swept volume and components such as the contact wire and catenary cable. Does it always exceed the preset safety threshold? To ensure electrical safety throughout the entire operation; Step 44: Perform segmented security assessment on the complete path list; if the intermediate sweep volume is detected to intrude into the constraint space described in Step 1, immediately execute the instruction interception logic, block the issuance permission of the path segment, and feed back the index and coordinate data of the interference node to the policy network in Step 2 in real time for path replanning; if the entire path passes the geometric verification, mark the path list as a safe and executable state, and use it as the final compliance benchmark for the actions of the physical execution mechanism. Step 5: Perform sequential execution of the virtual and real closed loop based on the verified discrete-time joint angle sequence; Step 51: Using the communication plugin provided by the digital twin simulation platform, the verified discrete time-series angle sequence is encapsulated into control messages in sequence; using a distributed publish-subscribe mechanism, the pose nodes are distributed to the physical articulated boom control system in stages; according to the instructions, the physical end drives the motor and hydraulic pump station to complete the joint actions in physical order; if a sudden collision risk or inconsistent instructions occur at the physical layer, the digital twin system immediately triggers the forced interrupt logic, cuts off the physical drive source and locks all joints; Step 52: Using encoders and pose sensing sensors deployed on the physical articulated arm, collect the actual generalized coordinates of each joint in real time. The data, including the end-effector pose, is fed back to the digital twin platform in real time via a communication link to drive the virtual model to reconstruct its pose. This involves using the PhysX engine to calculate the constraints between the virtual pose and the environment in real time, thereby achieving state synchronization and timing alignment between the virtual and real systems during execution.

[0066] Step 53: Calculate the preset pose nodes of the digital twin. Real-time pose feedback Error vector between ; If the deviation exceeds the preset threshold, the system will use a closed-loop control algorithm or feed the residual back to the policy network; the corrected joint target value will be calculated in real time using the dynamic feedback data of the PhysX engine, and the compensated increment will be added to the control signal of the currently active joint to achieve real-time correction of the physical end action. Step 54: After each round of sequential execution, the system determines whether the end effector has achieved the preset target through pose feedback; If the task is completed, the system generates a work completion beacon and drives the articulated boom to retract to its initial position along a preset path, completing the entire virtual-real closed-loop sequential work process.

[0067] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for sequential control of insulator wiping arm movement based on digital twin, characterized in that, Includes the following steps: Step 1: Construct a digital twin platform with constrained space; Step 2: Train a deep reinforcement learning policy network to generate motion instructions that adapt to the sequential execution pattern; Step 3: Generate discrete temporal joint angle sequences based on a deep reinforcement learning policy network; Step 4: Perform pose safety verification and collision detection on discrete temporal joint angle sequences based on the digital twin platform; Step 5: Perform sequential execution of the virtual and real closed loop based on the verified discrete time-series joint angle sequence.

2. The insulator wiping arm sequence control method based on digital twin according to claim 1, characterized in that, The specific process of step 1 includes: Step 11: Construct a digital articulated boom model in the digital twin platform based on the actual structural parameters of the boom; Step 12: Based on the structural parameters and working environment of the work area, construct a working environment model including obstacles and safety constraints in the digital twin platform; Step 13: Establish a communication mechanism based on the robot operating system between the digital twin platform and the physical articulated arm control system.

3. The insulator wiping arm sequence control method based on digital twin according to claim 2, characterized in that, The operational environment model includes a digital scene comprising maintenance vehicles, railway tracks, overhead contact line columns, overhead contact line cantilever arms, cantilever arm support pipes, positioners, contact wires, catenary wires, droppers, and high-voltage conductors.

4. The insulator wiping arm sequence control method based on digital twin according to claim 2, characterized in that, The operational environment model configures high-precision collision body attributes for each component. According to the electrified railway safety operation procedures, the boom and wiping tool must maintain a minimum safe distance from the contact line and catenary. At the same time, a safe distance constraint space is constructed by extending outward from the contact line as a reference. The constraint space is transformed into a collision penalty field in the physics simulation engine, so that the sweep volume of the boom body and the end wiping tool in the virtual environment can be monitored in real time, thereby providing accurate environmental geometric boundaries and safe clearance constraints for subsequent motion planning.

5. The insulator wiping arm sequence control method based on digital twin according to claim 1, characterized in that, The input to the deep reinforcement learning policy network includes a composite state observation vector. ; In the formula: The state vector of the curved arm body; It is a set of generalized joint coordinates; The rotation angle; and These are the position coordinates and attitude quaternions of the end effector in three-dimensional space, respectively; Guide the feature vector for the target; The coordinates of the center of the target insulator. The distance between the end effector and the target point is the Euclidean distance. The unit direction vector that guides the movement of the end effector; Features for sensing environmental obstacles; Here are the coordinates of the obstacle's position. For the first bent arm Link and the first The shortest distance between obstacles The minimum safe clearance between the end effector and the obstacle; This is a vector indicating the active state of a joint; This is the index number of the active joint currently determined based on timing logic.

6. The insulator wiping arm sequence control method based on digital twin according to claim 1, characterized in that, The deep reinforcement learning policy network constructs action mask logic at the policy inference layer, which is mathematically expressed as follows: In the formula: This is the index number of the active joint currently determined based on timing logic; This is the final execution instruction after masking correction; The original action vector output by the policy network; This is the action mask vector.

7. The insulator wiping arm sequence control method based on digital twin according to claim 1, characterized in that, The reward function of the deep reinforcement learning policy network Rewards from mission objectives Safety boundary penalties and dynamic smoothing penalty It consists of three parts, used to guide the SAC strategy model to achieve safe and accurate path search during parallel simulation; In the formula: This represents the negative exponential distance between the end effector and the target point; The Euclidean distance between the current position of the end effector and the target wiping point; The safety boundary penalty is designed with a distance for the overhead contact line cantilever and surrounding obstacles; This is the shortest distance between the edge of the swept volume and the contact line; This is the preset minimum safety gap threshold; This is a penalty term for a positive constant. The dynamic smoothing penalty is designed to address the sensitivity of electro-hydraulic hybrid drive systems to abrupt changes in motion. This represents the effective action quantity at the current moment after being corrected by the masking operator.

8. The insulator wiping arm sequence control method based on digital twin according to claim 1, characterized in that, The specific process of step 3 includes: Step 31: Solidify the parameters of the converged policy network during training to build an offline inference engine; Step 32: Obtain the three-dimensional coordinates of the insulator to be wiped in the robot coordinate system, and use them as the target guidance feature vector. The input reference is used; simultaneously, the physical poses of each joint of the articulated arm are collected as the initial state to initialize the composite state observation vector. The state vector of the curved arm body ; Step 33: The path planner inputs the current composite state observation vector into the fixed strategy model, uses the joint active state indicator vector in it to identify the motion stage, and iteratively calculates the target pose increment of the current active joint by combining the action masking mechanism. Based on this, it updates and generates a complete 7-dimensional joint space pose node. Then, it realizes the triggering of the next stage action and the construction of discrete sequence by automatically migrating the joint active state indicator bit. Step 34, through The process involves sequential, iterative reasoning rounds until all seven joints have completed their target searches, ultimately constructing a framework containing... A complete linked list of discrete job paths for an ordered sequence of nodes; Step 35: Perform a smoothness check and normalization inverse calculation on the generated angle sequence, and restore the output normalized values ​​to radians and displacement units that can be recognized by the physical layer.

9. The insulator wiping arm sequence control method based on digital twin according to claim 1, characterized in that, The specific process of step 4 includes: Step 41, generate the included The complete path list of each pose node is imported into the digital twin simulation platform, and the high-fidelity virtual articulated arm model is driven sequentially according to the ordered nodes in the list to restore the entire process trajectory of the physical entity from the initial pose to the target work position in the digital space. Step 42: For the single-axis progressive motion in each round of the curved arm, calculate the sweep volume formed by the motion envelope between adjacent pose nodes; Step 43: Monitor the geometric feature relationship between the swept volume and the obstacle constraint space constructed in the first step in real time, and verify the path and check the safety distance; Step 44: Perform segmented security assessment on the complete path list; if the intermediate sweep volume is detected to intrude into the constraint space described in Step 1, immediately execute the instruction interception logic, block the issuance permission of the path segment, and feed back the index and coordinate data of the interference node to the policy network in Step 2 in real time for path replanning; if the entire path passes the geometric verification, mark the path list as a safe and executable state, and use it as the final compliance benchmark for the actions of the physical execution mechanism.

10. The insulator wiping arm sequence control method based on digital twin according to claim 1, characterized in that, The specific process of step 5 includes: Step 51: Using the communication plugin provided by the digital twin simulation platform, the verified discrete time-series angle sequence is encapsulated into control messages in sequence; using a distributed publish-subscribe mechanism, the pose nodes are distributed to the physical articulated boom control system in stages; according to the instructions, the physical end drives the motor and hydraulic pump station to complete the joint actions in physical order; if a sudden collision risk or inconsistent instructions occur at the physical layer, the digital twin system immediately triggers the forced interrupt logic, cuts off the physical drive source and locks all joints; Step 52: Using encoders and pose sensing sensors deployed on the physical articulated arm, the actual generalized coordinates and end-effector poses of each joint are collected in real time; the data is fed back to the digital twin platform in real time through the communication link to drive the virtual model to reconstruct the pose. Step 53: Calculate the error vector between the preset pose node of the digital twin and the real-time pose feedback; If the deviation exceeds the preset threshold, the system will use a closed-loop control algorithm or feed the residual back to the strategy network; calculate the corrected joint target value in real time, and add the compensated increment to the control signal of the currently active joint to achieve real-time correction of the physical end action; Step 54: After each round of sequential execution, the system determines whether the end effector has achieved the preset target through pose feedback; If the task is completed, the system generates a work completion beacon and drives the articulated boom to retract to its initial position along a preset path, completing the entire virtual-real closed-loop sequential work process.

Images