Transformer detection mechanical arm control method and system based on hierarchical architecture

Abstract

The present application relates to the technical field of automation control, in particular to a transformer detection mechanical arm control method and system based on hierarchical architecture, which realizes centralized control and collaborative work of multiple mechanical arms by constructing three-layer architecture of task scheduling layer, path planning layer and action execution layer, and decomposing, distributing and transforming macroscopic detection tasks into specific mechanical arm action instructions step by step. The task scheduling layer is responsible for splitting macroscopic tasks into independent executable meta-tasks according to detection requirements, regulations and mechanical arm states, and distributing appropriate mechanical arms for each meta-task, thereby optimizing task distribution logic and avoiding resource idling or overload problems caused by fixed workstation configuration. The path planning layer obtains accurate task coordinates and attitude parameters for each meta-task through instruction semantic analysis and key feature extraction, and then generates collision-free and efficient motion trajectories, ensuring that the mechanical arm safely and accurately reaches the target position in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and in particular to a control method and system for a transformer detection robotic arm based on a hierarchical architecture. Background Technology

[0002] Transformer testing is a crucial step in the manufacturing of power equipment. Traditional testing methods primarily rely on single-station serial operation with manual assistance, resulting in low automation, long testing cycles, and low resource utilization. With the expansion of power equipment production capacity, testing methods are gradually shifting from low-frequency, single-unit testing to high-frequency, batch testing. The industry is increasingly introducing robotic arms to replace manual labor in repetitive tasks such as wiring, testing, and disconnection. Currently, typical robotic arm configurations employ a one-arm-per-station model, where each testing station is independently equipped with a robotic arm and auxiliary equipment. Their motion control is mostly based on single-machine sequential control or master-slave linkage control, lacking a multi-robotic arm centralized control technology framework. This lack of technology addresses issues such as overall line cycle time, resource deadlock, collision zone mutual exclusion, and asynchronous step compensation, hindering the ability to achieve multi-station operation with a single arm.

[0003] Specifically, existing transformer testing lines suffer from high hardware investment and maintenance costs due to their architecture, which uses independently configured robotic arms, controllers, and auxiliary equipment at each workstation. Each set of hardware requires independent calibration and maintenance, leading to inefficient maintenance as problems can only be addressed one by one in case of sudden failures. In actual operation, the utilization rate of robotic arms varies greatly across different workstations. Workstations performing rapid tests such as insulation ratio testing are often operating at full capacity, while other workstations may experience idle hardware resources, highlighting a significant resource imbalance. Furthermore, the existing robotic arm control methods cannot adapt to fluctuations in testing duration and cannot dynamically adjust based on the expected busy / idle status of each workstation. The rigid control timing logic settings result in a lack of synchronization between robotic arm status, task status, and workstation status. When faced with unexpected interruptions in testing due to instrument failures or power fluctuations, the system's response capability is poor, often requiring manual restoration of breakpoints and triggering of subsequent actions before the timing token can continue flowing from the breakpoint database, impacting the automated and continuous operation of the production line. In addition, the obstacle avoidance function of existing robotic arms mainly relies on physical isolation such as track barriers and safety light curtains. Their movement path is usually a fixed preset pattern. When encountering temporary obstacles such as tools being dropped or wires being tangled, the system cannot handle them autonomously and still needs to rely on manual intervention to clear the obstacles, which limits the intelligence level and operating efficiency of the inspection line. Summary of the Invention

[0004] This invention aims to provide a control method and system for a transformer testing robotic arm based on a hierarchical architecture, in order to construct a centralized control architecture for multiple robotic arms, and to solve the technical problems of high hardware investment and maintenance costs, prominent resource idleness, and low operating efficiency in the existing architecture where each workstation independently configures robotic arms, controllers, and auxiliary equipment.

[0005] To achieve the above objectives, the first aspect of the present invention provides a control method for a transformer detection robotic arm based on a hierarchical architecture, used to control a cluster of robotic arms. This method is implemented based on a preset hierarchical architecture, which includes a task scheduling layer, a path planning layer, and an action execution layer. The method includes the following steps: The task scheduling layer responds to the transformer detection command, obtains the macro task based on the transformer detection command, and then decomposes the macro task into several meta tasks according to the preset transformer detection procedure, the preset standard operating procedure, and the model and status of each robot arm in the robot arm cluster. Then, a robot arm is assigned to each meta task, and several meta tasks are sent to the path planning layer. For any given meta-task, the path planning layer obtains the transformer corresponding to the meta-task, performs instruction semantic parsing and key feature extraction on the meta-task based on the transformer and robotic arm corresponding to the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, and then generates the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and task posture parameters, and sends the motion trajectory to the action execution layer. The motion execution layer controls the movement of the corresponding robotic arm based on the motion trajectory, so that the robotic arm completes the corresponding meta-task.

[0006] The aforementioned layered architecture-based control method for transformer inspection robotic arms decomposes, allocates, and transforms macroscopic inspection tasks into specific robotic arm motion commands through a three-layer architecture consisting of a task scheduling layer, a path planning layer, and a motion execution layer. This enables centralized control and collaborative operation of multiple robotic arms. The task scheduling layer is responsible for breaking down macroscopic tasks into independently executable meta-tasks based on inspection requirements, procedures, and robotic arm status. It then assigns a suitable robotic arm to each meta-task, optimizing the task allocation logic and avoiding resource idleness or overload issues caused by fixed workstation configurations. The path planning layer, for each meta-task, obtains precise task coordinates and attitude parameters through instruction semantic parsing and key feature extraction, thereby generating a collision-free, high-efficiency motion trajectory to ensure the robotic arm safely and accurately reaches the target position in complex environments. The motion execution layer employs a PID control algorithm to adjust the motion state of each joint of the robotic arm in real time according to the planned trajectory, ensuring the accuracy of trajectory tracking and dynamic response speed. This layered architecture makes the responsibilities of each layer clear and the interfaces well-defined. It supports the parallel operation of multiple robotic arms, facilitates system maintenance and functional expansion, significantly reduces hardware investment and operation and maintenance costs, and improves the automation level and operating efficiency of the transformer testing line.

[0007] Further, the task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet; wherein, sending a plurality of the meta-tasks to the path planning layer includes: A task activation token is generated, and a meta-task repository is established for each meta-task based on the task activation token. Then, each meta-task repository is passed to the path planning initiation repository of the path planning Petri subnet through the corresponding substitution transition, triggering the calculation logic of the path planning Petri subnet. Sending the motion trajectory to the action execution layer includes: A path availability status token is generated, and then the path availability status token is returned to the task scheduling Petri subnet through the port library. The motion trajectory is also transmitted to the action execution Petri subnet through the port library, so that the task scheduling Petri subnet issues an execution token corresponding to the motion trajectory to the action execution Petri subnet based on the path availability status token.

[0008] In this implementation, the task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet. By leveraging the Petri net's transition and place mechanism, the system achieves modular decomposition of the layered architecture, event-driven interaction and timing control between layers, and avoids deadlock and collision risks through verification methods, thus ensuring reliable system operation.

[0009] It should be noted that Petri nets are a formal model that can graphically and accurately describe complex behaviors such as concurrency, synchronization, asynchrony, mutual exclusion, and conflict in software systems.

[0010] This system employs a Petri net model to formalize the interactions between task scheduling, path planning, and action execution layers as a token generation and transmission mechanism. The task scheduling layer generates a task activation token, establishes a placeholder for each meta-task, and transmits it to the path planning layer via substitution transitions, triggering path planning calculations. Upon completion, the path planning layer generates a path availability status token, simultaneously returning the status token to the task scheduling layer and transmitting the motion trajectory to the action execution layer via the port placeholder. The task scheduling layer then issues execution tokens to the action execution layer based on this information. This mechanism leverages the concurrency, synchronization, and resource-sharing capabilities of Petri nets to achieve asynchronous communication and collaborative control among the three layers. It effectively solves deadlock, resource contention, and cycle time mismatch problems in multi-robotic arm operations, enhancing the system's fault tolerance and real-time response capabilities, and ensuring the continuous and reliable execution of detection tasks.

[0011] Further, the task scheduling layer responds to the transformer detection command, obtains the macro-task based on the transformer detection command, and then decomposes the macro-task into several meta-tasks according to the preset transformer detection procedure, the preset standard operating procedure, and the model and status of each robotic arm in the robotic arm cluster. Then, a robotic arm is assigned to each meta-task, and the several meta-tasks are sent to the path planning layer, including: Obtain the logical dependencies and execution priorities between each of the meta-tasks, and then determine the task execution order of several meta-tasks based on the logical dependencies and execution priorities; In response to any abnormal execution result of the meta-task, the emergency response plan task corresponding to the abnormal execution result is obtained, and then the task execution order is updated based on the emergency response plan task; A robotic arm is assigned to each of the meta-tasks according to the task execution order, and several of the meta-tasks are sent to the path planning layer.

[0012] In this implementation, by introducing task dependencies, priority sorting, and dynamic emergency response mechanisms, the task scheduling layer can reasonably arrange the execution order of meta-tasks according to the constraints of the actual testing process, avoiding waiting or conflicts caused by improper task order. At the same time, when abnormal events such as instrument failure or power fluctuations occur during the testing process, the system can quickly call the preset emergency plan tasks, dynamically adjust the execution order of the remaining tasks, and prioritize operations such as abnormal recovery or safe shutdown, thereby greatly improving the production line's autonomous ability to cope with emergencies and the stability of continuous operation, and reducing manual intervention and downtime.

[0013] Further, the step of obtaining the transformer corresponding to the meta-task, and then performing instruction semantic parsing and key feature extraction on the meta-task based on the transformer and the robotic arm corresponding to the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, includes: Obtain the 3D digital model of the transformer corresponding to the meta-task, and the 3D digital model of the robotic arm corresponding to the meta-task. The instruction semantics of the meta-task are parsed to obtain the core actions, detection objects, and task constraints corresponding to the meta-task. Key features are extracted from the meta-task to obtain the target position features, end effector attitude features, and motion parameter features corresponding to the meta-task. Based on the three-dimensional digital model of the transformer, the three-dimensional digital model of the robotic arm, the core motion, the detection object, the task constraints, the target position features, the end effector posture features, and the motion parameter features, model matching and coordinate transformation are performed to calculate the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task.

[0014] In this implementation, three-dimensional digital models and semantic parsing technology are used to transform abstract meta-task instructions into specific geometric parameters. This enables the path planning layer to accurately understand the task intent and automatically adapt to the size and structural differences of different transformer models and robotic arms. No manual teaching or repeated debugging is required, which significantly improves the versatility and accuracy of path planning and reduces the deployment cycle and technical threshold.

[0015] Further, generating the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and the task posture parameters includes: The path start point and path end point are obtained based on the task coordinate parameters and the task attitude parameters; Based on the global search algorithm and the fast expanding random tree algorithm, the geometric path from the starting point of the path to the ending point of the path is obtained; The dynamic model and kinematic constraints of the robotic arm are obtained, and then the geometric path is interpolated and optimized based on the dynamic model and kinematic constraints to generate a motion trajectory.

[0016] In this implementation, by integrating a global search algorithm with a fast expanding random tree algorithm, the method can quickly search for collision-free geometric paths while ensuring global optimality. This solves the problem that single algorithms are prone to getting trapped in local optima or having low search efficiency in complex environments. Furthermore, by combining the dynamic model and kinematic constraints of the robotic arm, the geometric path is interpolated and optimized to ensure that the generated motion trajectory meets physical constraints such as joint velocity and acceleration. This ensures that the robotic arm can perform tasks smoothly and stably, extends equipment life, and improves operational safety.

[0017] Furthermore, obtaining the geometric path from the starting point to the ending point of the path based on the global search algorithm and the fast expanding random tree algorithm includes: Obtain the comprehensive joint motion cost and obstacle safety cost of the corresponding robotic arm, and then construct the actual cost function based on the comprehensive joint motion cost and the obstacle safety cost; Obtain the joint angle difference and end-effector position difference corresponding to the robotic arm, and then construct a heuristic cost function based on the joint angle difference and the end-effector position difference; A total cost function is constructed based on the actual cost function and the heuristic cost function, and joint angular velocity constraints and acceleration constraints are added to the total cost function. Based on the total cost function, the path start point, and the path end point, a path iterative search is performed in the preset configuration space of the robotic arm to obtain the geometric path from the path start point to the path end point.

[0018] In this implementation, by constructing an actual cost function that includes joint motion costs and obstacle safety costs, as well as a heuristic cost function based on joint angle differences and end-effector position differences, this method considers both the economy and safety of motion during path search, and introduces goal-oriented heuristic information, effectively improving search efficiency and path quality. At the same time, by incorporating joint angular velocity and acceleration constraints into the total cost function, it ensures that the searched geometric path can meet the dynamic feasibility requirements after subsequent interpolation, thereby improving the reliability and practicality of path planning.

[0019] Furthermore, the step of controlling the motion execution layer to move the corresponding robotic arm based on the motion trajectory, so that the robotic arm completes the corresponding meta-task, includes: Based on the motion trajectory, the desired angle, desired angular velocity, and desired angular acceleration of each joint of the robotic arm are obtained; The actual angle, actual angular velocity, and actual angular acceleration of each joint of the robotic arm are detected, and then the angle error between the desired angle and the actual angle, the angular velocity error between the desired angular velocity and the actual angular velocity, and the angular acceleration error between the desired angular acceleration and the actual angular acceleration are calculated respectively. Based on the angle error, the angular velocity error, and the angular acceleration error, a joint control signal is generated using a PID control algorithm. Then, the movement of the corresponding robotic arm is controlled according to the joint control signal so that the robotic arm can complete the corresponding meta-task.

[0020] In this implementation, a multi-closed-loop PID control strategy based on angle, angular velocity, and angular acceleration errors is adopted, which can compensate for nonlinear disturbances and modeling errors in the movement of the robotic arm in real time, ensuring that each joint accurately tracks the desired trajectory. This is especially suitable for operation scenarios in transformer detection that require high-precision positioning and stable contact. By simultaneously introducing angular velocity and angular acceleration feedback, oscillations and overshoots during the movement are effectively suppressed, improving the dynamic response performance and robustness of the system and ensuring the quality of the completion of the meta-task.

[0021] A second aspect of this invention provides a layered architecture-based control system for a transformer detection robotic arm, used to control a cluster of robotic arms, comprising a task scheduling layer, a path planning layer, and an action execution layer; wherein: The task scheduling layer is used to obtain transformer detection instructions, obtain macro tasks based on the transformer detection instructions, and then decompose the macro tasks into several meta tasks according to the preset transformer detection procedures, preset standard operating procedures and the model and status of each robotic arm in the robotic arm cluster. Then, a robotic arm is assigned to each meta task, and the several meta tasks are sent to the path planning layer. The path planning layer is used for any of the meta-tasks: obtaining the transformer corresponding to the meta-task, and based on the transformer and the robotic arm corresponding to the meta-task, performing instruction semantic parsing and key feature extraction on the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, and then generating the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and the task posture parameters, and sending the motion trajectory to the action execution layer. The motion execution layer is used to control the movement of the corresponding robotic arm based on the motion trajectory using a PID control algorithm, so that the robotic arm can complete the corresponding meta-task.

[0022] Further, the task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet; wherein, sending a plurality of the meta-tasks to the path planning layer includes: A task activation token is generated, and a meta-task repository is established for each meta-task based on the task activation token. Then, each meta-task repository is passed to the path planning initiation repository of the path planning Petri subnet through the corresponding substitution transition, triggering the calculation logic of the path planning Petri subnet. Sending the motion trajectory to the action execution layer includes: A path availability status token is generated, and then the path availability status token is returned to the task scheduling Petri subnet through the port library. The motion trajectory is also transmitted to the action execution Petri subnet through the port library, so that the task scheduling Petri subnet issues an execution token corresponding to the motion trajectory to the action execution Petri subnet based on the path availability status token.

[0023] Furthermore, the task scheduling layer is used to obtain transformer detection instructions, obtain macro-level tasks based on the transformer detection instructions, and then decompose the macro-level tasks into several meta-tasks according to the preset transformer detection procedures, preset standard operating procedures, and the model and status of each robotic arm in the robotic arm cluster. Then, a robotic arm is assigned to each meta-task, and the several meta-tasks are sent to the path planning layer, including: Obtain the logical dependencies and execution priorities between each of the meta-tasks, and then determine the task execution order of several meta-tasks based on the logical dependencies and execution priorities; In response to any abnormal execution result of the meta-task, the emergency response plan task corresponding to the abnormal execution result is obtained, and then the task execution order is updated based on the emergency response plan task; A robotic arm is assigned to each of the meta-tasks according to the task execution order, and several of the meta-tasks are sent to the path planning layer. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a control method for a transformer detection robotic arm based on a hierarchical architecture, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall control flow of a layered architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Petri subnet library, token, and timing control structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the task priority and exception handling process of the task scheduling layer provided in the embodiment of the present invention; Figure 5 This is a schematic diagram of a transformer detection robotic arm control system based on a hierarchical architecture, provided in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the following detailed descriptions are exemplary and intended to provide further detailed explanation of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0026] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0027] It should be noted that current transformer testing methods are mainly single-station serial and manually assisted, resulting in low automation, long testing cycles, and low resource utilization. With the expansion of power equipment production capacity, testing methods are shifting from single-unit, low-frequency to batch, high-frequency operations. The industry is gradually introducing robotic arms to replace manual labor in repetitive tasks such as wiring, testing, and disconnection. A typical robotic arm configuration adopts a "one arm per station" model, with each testing station independently equipped with one robotic arm and auxiliary equipment. The robotic arm motion control is a single-machine-level sequential control or master-slave linkage control, technically based on a simple mode of preset actions. It lacks a system-level, cycle-level multi-robotic arm centralized control technology framework, and is deficient in technologies related to overall line cycle time, resource deadlock, collision zone mutual exclusion, and asynchronous step compensation, thus failing to achieve the flexible "one arm, multiple stations" operation capability.

[0028] Based on this, the existing transformer testing methods have at least the following drawbacks: Firstly, the robotic arm hardware configuration is redundant, resulting in significant idle operating resources. The existing transformer testing line uses independently configured robotic arms, controllers, and auxiliary equipment for each workstation. Each set of hardware requires independent calibration and maintenance, meaning that any sudden malfunction must be investigated piecemeal, leading to high hardware investment and maintenance costs. The utilization rate of robotic arms varies greatly across different workstations; slow-speed tests such as temperature rise and withstand voltage tests have an average utilization rate of less than 20%, while fast-speed tests such as insulation ratio tests often operate at full load, highlighting the problem of idle hardware resources.

[0029] Secondly, the robotic arm control timing is rigid, lacking the ability to resume actions from interruptions. The robotic arm control cannot adapt to fluctuations in detection time or adjust according to the expected busy / idle status of the workstation. The rigid control timing logic settings result in a lack of synchronization between the robotic arm state, task state, and workstation state, leading to poor handling of unexpected events. In the event of unexpected interruptions such as instrument malfunctions or power fluctuations, manual restoration of the breakpoint and triggering of subsequent actions are required for the timing token to continue flowing from the breakpoint library.

[0030] Third, path conflicts depend on physical isolation, resulting in a lag in the robotic arm's obstacle avoidance response.

[0031] Current robotic arms rely on track barriers and safety light curtains for obstacle avoidance to achieve physical isolation. The robotic arm's movement path is fixed and preset; when encountering temporary obstacles such as dropped tools or tangled wires, manual intervention is required to clear them, with an average processing time exceeding 10 minutes. The low sampling frequency of position data and poor accuracy in predicting collision risks at path intersections increase the risk of collisions with transformer casings.

[0032] To address the aforementioned technical problems, the following embodiments of the present invention first construct a timing network for centralized control of multiple robotic arms. The robotic arms possess flexible cross-workstation collaboration capabilities, breaking through the original independent control layout of "one arm per workstation." The prototype experimental line requires only half the hardware investment of the original configuration to achieve the same detection capabilities, thus solving the problem of idle hardware resources.

[0033] Secondly, the following embodiments of the present invention establish a hierarchical time-series coordination mechanism, wherein the Task Scheduling Layer (TPN) allocates detection tasks, the Path Planning Layer (PPN) plans workstation switching paths, and the Action Execution Layer (APN) controls the fine-grained operation of the end point, thereby realizing the time-series coordination of "task-path-action" and solving the problems of temporary time-series adjustment and breakpoint continuation.

[0034] Finally, the following embodiments of the present invention establish a technical framework for hierarchical control architecture and multi-dimensional perception fusion. Through a multi-modal fusion perception system, the physically isolated environmental cognition basis is replaced, realizing core capabilities such as visual sensor terminal positioning, force sensor wiring pressure adjustment, and position sensor joint encoder, thus solving a series of obstacle avoidance problems for robotic arms.

[0035] For details, please refer to Figure 1 The first embodiment of the present invention provides a control method for a transformer detection robotic arm based on a hierarchical architecture, used to control a cluster of robotic arms. This method is implemented based on a preset hierarchical architecture, which includes a task scheduling layer, a path planning layer, and an action execution layer. The method includes the following steps: The task scheduling layer responds to the transformer detection command, obtains the macro task based on the transformer detection command, and then decomposes the macro task into several meta tasks according to the preset transformer detection procedure, the preset standard operating procedure, and the model and status of each robot arm in the robot arm cluster. Then, a robot arm is assigned to each meta task, and several meta tasks are sent to the path planning layer. For any given meta-task, the path planning layer obtains the transformer corresponding to the meta-task, performs instruction semantic parsing and key feature extraction on the meta-task based on the transformer and robotic arm corresponding to the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, and then generates the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and task posture parameters, and sends the motion trajectory to the action execution layer. The motion execution layer controls the movement of the corresponding robotic arm based on the motion trajectory, so that the robotic arm completes the corresponding meta-task.

[0036] The aforementioned robotic arm control system employs a hierarchical control architecture to address the complex, hazardous, and highly precision-critical transformer inspection environment. This hierarchical architecture consists of a Task Scheduling Network (TPN), a Path Planning Network (PPN), and an Action Execution Network (APN) working collaboratively. The TPN decomposes the macroscopic inspection target into ordered, executable subtasks; the PPN transforms abstract tasks into concrete, collision-free motion trajectories; and the APN executes the trajectories accurately and in real-time, collecting sensor data. These three layers form a closed loop through top-down command transmission and bottom-up status feedback, ensuring the system can autonomously, efficiently, and safely complete the inspection task.

[0037] like Figure 2 As shown, in one specific embodiment, the transformer inspection robotic arm control system based on a hierarchical architecture further refines the functional modules of each layer and their interaction relationships. This architecture consists of a Task Scheduling Layer (TPN), a Path Planning Layer (PPN), and an Action Execution Layer (APN). The layers form a closed-loop control system through a top-down command flow and a bottom-up feedback flow, ensuring efficient, accurate, and safe execution of the inspection task.

[0038] The Task Scheduling Layer (TPN) is responsible for the decomposition and resource allocation of macro-level tasks. When the system receives a new detection task, it first performs semantic understanding of the transformer detection instructions through the task parsing module to form a macro-level task description. Subsequently, based on preset procedures such as transformer detection regulations and standard operating procedures, as well as the current model and status of the robotic arm, the meta-task decomposition module breaks down the macro-level task into a series of ordered and independently executable meta-tasks, generating an encodeable sequence, such as meta-task #1 to meta-task #N. The task scheduling module coordinates the execution order of the meta-tasks, while the resource scheduling module allocates each meta-task to a specific robotic arm in the robot group based on the availability of the robotic arm. Finally, the allocated meta-tasks are sent down to the path planning layer.

[0039] The Path Planning Layer (PPN) is responsible for transforming meta-tasks into specific motion trajectories. For each meta-task, this layer first obtains the geometric and dynamic parameters of the transformer and robotic arm through a physical model query, while simultaneously using the current parameter capture module to collect real-time environmental data, such as obstacle positions and robotic arm pose. The emergency path planning module can dynamically adjust its planning strategy when it detects sudden obstacles or path conflicts. The path algorithm module generates collision-free motion trajectories based on the meta-task requirements, the physical model, and real-time data. For different work scenarios, path planning is divided into two modes: cross-workstation and intra-workstation. For cross-workstation scenarios, discrete waypoints are generated and converted into trajectory instructions; for intra-workstation scenarios, continuous curves are generated. The final generated motion trajectory is sent to the motion execution layer.

[0040] The Action Execution Layer (APN) is responsible for precise trajectory tracking and real-time control of the robotic arm. Based on an embedded system and motion control card, this layer parses the received motion trajectory into desired motion parameters. Sensors collect actual motion parameters of the robotic arm in real time, such as joint angles, speed, and torque, and compare these with the desired values. The control algorithm module employs various strategies, including PID control, feedforward control, and adaptive control, to generate precise motion commands to drive the robotic arm's movement. Simultaneously, feedback information such as actual position and status is uploaded in real-time to the path planning layer and task scheduling layer for closed-loop adjustment and status monitoring, ensuring high-precision completion of the detection task.

[0041] The entire layered architecture achieves decoupling of task decomposition, path planning and action execution through modular design. The modules within each layer work together, and the instructions and feedback between layers are seamlessly connected, which fully adapts to the complexity and high precision requirements of the transformer testing environment.

[0042] The aforementioned layered architecture-based control method for transformer inspection robotic arms decomposes, allocates, and transforms macroscopic inspection tasks into specific robotic arm motion commands through a three-layer architecture consisting of a task scheduling layer, a path planning layer, and a motion execution layer. This enables centralized control and collaborative operation of multiple robotic arms. The task scheduling layer is responsible for breaking down macroscopic tasks into independently executable meta-tasks based on inspection requirements, procedures, and robotic arm status. It then assigns a suitable robotic arm to each meta-task, optimizing the task allocation logic and avoiding resource idleness or overload issues caused by fixed workstation configurations. The path planning layer, for each meta-task, obtains precise task coordinates and attitude parameters through instruction semantic parsing and key feature extraction, thereby generating a collision-free, high-efficiency motion trajectory to ensure the robotic arm safely and accurately reaches the target position in complex environments. The motion execution layer employs a PID control algorithm to adjust the motion state of each joint of the robotic arm in real time according to the planned trajectory, ensuring the accuracy of trajectory tracking and dynamic response speed. This layered architecture makes the responsibilities of each layer clear and the interfaces well-defined. It supports the parallel operation of multiple robotic arms, facilitates system maintenance and functional expansion, significantly reduces hardware investment and operation and maintenance costs, and improves the automation level and operating efficiency of the transformer testing line.

[0043] Further, the task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet; wherein, sending a plurality of the meta-tasks to the path planning layer includes: A task activation token is generated, and a meta-task repository is established for each meta-task based on the task activation token. Then, each meta-task repository is passed to the path planning initiation repository of the path planning Petri subnet through the corresponding substitution transition, triggering the calculation logic of the path planning Petri subnet. Sending the motion trajectory to the action execution layer includes: A path availability status token is generated, and then the path availability status token is returned to the task scheduling Petri subnet through the port library. The motion trajectory is also transmitted to the action execution Petri subnet through the port library, so that the task scheduling Petri subnet issues an execution token corresponding to the motion trajectory to the action execution Petri subnet based on the path availability status token.

[0044] In one specific embodiment, the hierarchical Petri net achieves modular decomposition of the transformer testing system through substitution transitions and hierarchical subnets, adapting to the complex requirements of multi-component transformer testing. The Task Scheduling Layer (TPN) breaks down the macroscopic objective of transformer testing into several meta-task places, corresponding to the testing requirements of each core component of the transformer, covering testing items for key components such as the tank, bushings, windings, and insulating oil. Each task place establishes an association with the corresponding subnet of the Path Planning Layer (PPN) through independent substitution transitions, achieving precise decomposition of macroscopic tasks into mid-level executable logic. Each substitution transition is mapped to a dedicated path planning subnet of the PPN. This subnet further decomposes the task into motion path parameters for the robotic arm based on the structural characteristics of the corresponding transformer component, including core information such as motion trajectory, target pose, and speed limits. The path planning results of the PPN are transmitted to the action execution subnet of the Action Execution Layer (APN) through port places. This subnet converts the path parameters into control signals for the servo motors of the robotic arm joints, driving the robotic arm to carry the testing instruments to complete the testing action, realizing a closed loop from path planning to physical execution.

[0045] The system utilizes hierarchical Petri nets with synchronized transitions and shared storage to achieve event-driven data interaction between layers, ensuring orderly collaboration in the transformer testing process. Upon receiving an external testing task, the task storage in the TPN layer generates a task activation token, which is then passed to the corresponding path planning initiation storage in the PPN layer via a substitution transition. This triggers the transition sequence of the path planning subnet and initiates the path calculation logic, ensuring rapid task response from the top to the middle layers. After completing path planning, the PPN layer returns a path availability status token to the TPN layer through the output port storage. The TPN layer then sends an action execution command to the APN layer based on this token, activating the APN layer's action execution subnet and achieving execution linkage between the middle and bottom layers. Real-time data from the joint encoders, force sensors, and various testing instruments mounted on the robotic arm are continuously fed back to the action execution layer (APN) through the sensor storage. When deviations in the robotic arm's motion trajectory or abnormal contact pressure are detected, the APN layer immediately triggers a compensation transition to adjust the motion parameters and notifies the PPN layer to update the path through the port storage, ensuring that the testing actions always meet accuracy requirements and guaranteeing the accuracy of the testing data.

[0046] The system employs reachability analysis and invariant verification using hierarchical Petri nets to ensure the logical correctness of the transformer detection control architecture, thereby improving the system's reliability and security. Analysis of the hierarchical network's reachability graph identifies potential deadlock states during transformer detection, and pre-defined deadlock prevention strategies eliminate deadlock risks, ensuring the continuous operation of the detection process. Invariant verification is performed on the path planning of the PPN layer, specifically through judgment equations. The solvability is verified by invariant, where... For the place vector, The incidence matrix, the matrix Position elements are transitions Triggering a database The token change amount. Through invariant verification, it is ensured that the robotic arm's movement path always maintains a safe distance from the transformer body, detection line fixing facilities, etc., to avoid collisions that could cause equipment damage or detection interruption.

[0047] Please refer to Figure 3 ,exist Figure 3 As shown in the Petri net diagram illustrating the token and timing control structure, the transformer detection robotic arm control system based on a hierarchical architecture further employs a hierarchical Petri net to achieve logical modeling within each layer and interaction between layers. This Petri net model consists of a Task Scheduling Petri Net (TPN), a Path Planning Petri Net (PPN), and an Action Execution Petri Net (APN). Through token passing and transition triggering mechanisms, it achieves layer-by-layer decomposition, path planning, and precise execution of the detection task.

[0048] At the Task Scheduling Network (TPN), the system takes the detection target as input and uses reachability graph analysis to prevent deadlocks in the task scheduling logic, ensuring that the macro-level task decomposition process is non-blocking. Each meta-task corresponding to the decomposed task library is associated with the path planning layer through substitution transitions #1 to #N. When an external detection task arrives, the TPN generates a task activation token, activates the corresponding meta-task library, and passes the token to the path planning initiation library in the PPN layer through substitution transitions, triggering the computation logic of the path planning subnet.

[0049] At the Path Planning Layer (PPN), upon receiving the task token, the system initiates a transition sequence to generate specific motion parameters. Motion parameters #1 to #N in the diagram represent core information such as the planned robotic arm trajectory, target pose, and speed limits. Simultaneously, the PPN, through an invariant verification module, ensures the planned path maintains a safe distance from the transformer and surrounding facilities, avoiding collisions, based on the solvability of the correlation matrix equation AX=0. After planning is complete, the PPN returns a status token (indicating path readiness) to the TPN via the port library and transmits the motion trajectory to the APN via the same port library. Furthermore, the PPN includes a compensation transition #N, used to receive feedback from the APN and dynamically adjust motion parameters to adapt to environmental changes.

[0050] At the Action Execution Layer (APN), after receiving the motion trajectory from the PPN, the system converts it into control signals #1 to #N, driving servo motors #1 to #N to execute specific actions. The APN has a built-in sensing library that collects data in real-time from the robotic arm joint encoders, force sensors, and detection instruments, forming execution feedback. When trajectory deviation or abnormal contact pressure is detected, the APN triggers a fault-solving mechanism, notifying the PPN to update the path through compensation transitions to ensure detection accuracy. Simultaneously, the sensing data is fed back upwards through the port library, forming a closed-loop control system.

[0051] The entire hierarchical Petri net maps macroscopic tasks to mid-level path planning through substitution transitions, facilitates bidirectional transmission of instructions and states between layers through a port library, and ensures the logical correctness and operational security of the system through reachability graph analysis and invariant verification. This model is fully adaptable to the complex requirements of multi-component transformer detection, achieving full-process controllability and traceability from task decomposition to action execution.

[0052] In this implementation, the task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet. By leveraging the Petri net's transition and place mechanism, the system achieves modular decomposition of the layered architecture, event-driven interaction and timing control between layers, and avoids deadlock and collision risks through verification methods, thus ensuring reliable system operation.

[0053] It should be noted that Petri nets are a formal model that can graphically and accurately describe complex behaviors such as concurrency, synchronization, asynchrony, mutual exclusion, and conflict in software systems.

[0054] This system employs a Petri net model to formalize the interactions between task scheduling, path planning, and action execution layers as a token generation and transmission mechanism. The task scheduling layer generates a task activation token, establishes a placeholder for each meta-task, and transmits it to the path planning layer via substitution transitions, triggering path planning calculations. Upon completion, the path planning layer generates a path availability status token, simultaneously returning the status token to the task scheduling layer and transmitting the motion trajectory to the action execution layer via the port placeholder. The task scheduling layer then issues execution tokens to the action execution layer based on this information. This mechanism leverages the concurrency, synchronization, and resource-sharing capabilities of Petri nets to achieve asynchronous communication and collaborative control among the three layers. It effectively solves deadlock, resource contention, and cycle time mismatch problems in multi-robotic arm operations, enhancing the system's fault tolerance and real-time response capabilities, and ensuring the continuous and reliable execution of detection tasks.

[0055] It should be understood that the Task Scheduling Layer (TPN) is the core of the entire system's intelligent decision-making. Located at the highest level of the control architecture, it is responsible for macro-level task management and scheduling, ensuring that the entire detection process can proceed efficiently and orderly according to predetermined goals, procedures, and priorities. The main responsibility of the TPN is to receive abstract task instructions from operators, transform them into a series of specific, executable meta-task sequences, and then send them down to the lower Path Planning Layer (PPN) for execution.

[0056] Therefore, the task scheduling layer responds to transformer detection commands, obtains macro-level tasks based on these commands, and then decomposes these macro-level tasks into several meta-tasks according to preset transformer detection procedures, preset standard operating procedures, and the model and status of each robotic arm in the robotic arm cluster. In one specific embodiment, the TPN serves as the system entry point, receiving transformer detection commands from external sources. These detection requirements come from transformer manufacturers, power grid companies, or power engineering companies. The TPN possesses the ability to parse structured commands, accepts mainstream data formats in the power industry, and accurately understands the final goal, constraints, and scope of equipment involved in the task. The next step is task decomposition. After understanding the macro-level task, the TPN decomposes it into a series of meta-tasks based on the built-in transformer detection procedures, standard operating procedures (SOPs), and the specific model and status of the equipment.

[0057] For example, in one possible embodiment, the transformer factory testing tasks can be decomposed into 12 primary testing items, totaling 38 specific meta-tasks, covering all key items of the transformer factory test. The specific decomposition is as follows: Macroscopic tasks for appearance and structural dimension inspection include four sub-tasks. Among them, the specific sub-task for the appearance inspection of the fuel tank is used to check for welding defects, coating damage, and accessory installation problems; the sub-task for the appearance and installation dimensions inspection of the bushing is used to verify the integrity of the insulation components and the installation positioning accuracy; the sub-task for the appearance inspection of the radiator and pipeline is used to ensure that the heat dissipation components are free from deformation and the pipeline connections are reliable; and the sub-task for the precise measurement of the overall structural dimensions is used to verify that the overall shape of the equipment and the spacing between components meet the design requirements.

[0058] The macroscopic task of winding DC resistance testing includes four sub-tasks. The pre-test preparation and wiring verification sub-task is used to ensure the reliability of the test circuit and eliminate contact resistance interference; the tap changer resistance test sub-task is used to check the contact performance of the tap changer and the reliability of the winding connection; the resistance data temperature conversion and comparison sub-task is used to eliminate the influence of temperature and accurately determine the winding condition; and the post-test demagnetization treatment sub-task is used to avoid residual magnetism in the winding affecting the accuracy of subsequent tests.

[0059] Macroscopic tasks for core insulation resistance and grounding current testing: These are divided into four sub-tasks. The core overall insulation resistance test sub-task is used to evaluate the insulation performance between core laminations; the core-clamp insulation test sub-task is used to investigate potential short circuits between the core and metal parts; the grounding lead connection reliability test sub-task is used to ensure the mechanical strength and conductivity of the grounding system; and the no-load grounding current monitoring sub-task is used to verify the absence of multi-point grounding faults.

[0060] Macroscopic task of winding insulation resistance and absorption ratio test: includes 3 meta-tasks. Among them, the environmental condition calibration meta-task is used to eliminate the influence of temperature and humidity on insulation test; the high / medium / low voltage winding insulation resistance test meta-task is used to evaluate the insulation strength of each winding respectively; and the insulation resistance temperature conversion meta-task is used to achieve the comparability of data at different temperatures.

[0061] The macroscopic task of dielectric loss factor (tgδ) testing includes three meta-tasks: the winding tgδ test (20℃) meta-task is used to detect the degree of moisture and aging of the winding insulation dielectric; the bushing tgδ and capacitance test meta-task is used to evaluate the bushing insulation performance and capacitance stability; and the temperature compensation and data correction meta-task is used to ensure the accuracy of test results at different temperatures.

[0062] Macroscopic tasks for fuel tank sealing performance testing include four meta-tasks. The sealing test preparation meta-task is used to ensure the reliability of the test system and eliminate interference factors; the hydraulic sealing test meta-task is used to verify the sealing performance of the fuel tank under positive pressure; the vacuum sealing test meta-task is used to verify the sealing reliability under negative pressure; and the leak point location detection meta-task is used to accurately locate minute leak defects.

[0063] The macro-level tasks of cooling system performance testing include three meta-tasks: the fan and oil pump start-stop test meta-task is used to check the operational flexibility of the cooling components; the oil circuit circulation sealing test meta-task is used to ensure that there are no leaks in the oil circuit and that the circulation is smooth; and the rated load temperature rise test meta-task is used to verify that the cooling capacity meets the thermal balance requirements.

[0064] Macro-level tasks of tap changer switching test: Includes 3 sub-tasks. Among them, the sub-task of switching mechanism appearance and insulation inspection is used to check for mechanical structural defects and insulation hazards; the sub-task of manual / electric switching performance test is used to verify switching reliability and response speed; and the sub-task of tap changer oil chamber sealing test is used to ensure that there is no leakage in the oil chamber and that the insulating oil is not contaminated.

[0065] Macroscopic tasks of partial discharge testing (under rated voltage): include 3 meta-tasks. Among them, the test system calibration meta-task is used to ensure that the detection accuracy meets the requirements; the winding partial discharge detection meta-task is used to check for insulation defects inside the winding; and the insulating oil partial discharge detection meta-task is used to detect discharges caused by air bubbles or impurities in the oil.

[0066] Macroscopic tasks for short-circuit impedance testing include two meta-tasks: the test wiring and parameter setting meta-task is used to ensure the test circuit is correct and the instrument parameters are matched; the impedance value measurement and repeatability verification meta-task is used to verify the winding structure stiffness and measurement reliability.

[0067] Macroscopic tasks for no-load loss and no-load current testing include two meta-tasks: the no-load pressurization and data acquisition meta-task is used to evaluate the core magnetic circuit performance; the loss data correction and comparison meta-task is used to eliminate the influence of voltage deviation and accurately evaluate the core quality.

[0068] Macroscopic tasks for insulating oil performance testing include three meta-tasks: oil sample collection and pretreatment to ensure the representativeness and uncontaminated nature of the oil sample; breakdown voltage testing to assess the insulation strength of the oil sample; and moisture and dielectric loss testing to detect the degree of oil deterioration.

[0069] Furthermore, the task scheduling layer responds to the transformer detection command, obtains the macro-task based on the transformer detection command, and then decomposes the macro-task into several meta-tasks according to the preset transformer detection procedure, the preset standard operating procedure, and the model and status of each robotic arm in the robotic arm cluster. It then assigns a robotic arm to each meta-task and sends the meta-tasks to the path planning layer, including: Obtain the logical dependencies and execution priorities between each of the meta-tasks, and then determine the task execution order of several meta-tasks based on the logical dependencies and execution priorities; In response to any abnormal execution result of the meta-task, the emergency response plan task corresponding to the abnormal execution result is obtained, and then the task execution order is updated based on the emergency response plan task; A robotic arm is assigned to each of the meta-tasks according to the task execution order, and several of the meta-tasks are sent to the path planning layer.

[0070] In one specific embodiment, after the meta-tasks are decomposed and generated, they are managed for priority and exception handling by the scheduling module of the task scheduling layer. The task scheduling layer generates a real-time task dependency graph based on the logical dependencies and execution priorities between meta-tasks, assigns priorities to different meta-tasks, and determines the task execution order. The task scheduling layer employs a fixed contingency plan for exception handling. If any sub-task returns an abnormal result during the detection process, the task scheduling layer queries the contingency plan, temporarily inserts a high-priority task, and suspends other non-urgent tasks.

[0071] Please refer to Figure 4 ,exist Figure 4 The task scheduling layer's task priority and exception handling flowchart illustrates how the internal task scheduling layer's task priority scheduling and exception handling processes utilize state machines and decision nodes to achieve the orderly execution and dynamic adjustment of meta-tasks. This process encompasses two main threads: normal scheduling and emergency handling, ensuring that transformer testing tasks can proceed efficiently and reliably even in complex environments.

[0072] The process begins with the set of meta-tasks generated by the meta-task decomposition module. Each meta-task is first added to the waiting queue and enters a pending state. Subsequently, the system triggers the dependency analysis phase, determining whether the current meta-task meets its prerequisites based on the preset transformer detection procedures and the robotic arm status, such as whether the prerequisite tasks are completed and whether resources are ready. If the dependency conditions are met, the system enters the priority allocation phase, assigning appropriate priorities based on the urgency and logical relationships of the tasks; if not, the system switches to the exception reporting channel, triggering the exception handling mechanism.

[0073] Metatasks that pass dependency checks are added to the task pool and moved to the trigger queue after priority assignment, awaiting scheduling and execution. The priority scheduling module dynamically selects tasks to be executed based on task priority and trigger queue status, and generates execution instructions to be sent to the path planning layer. After task execution, the system receives feedback signals from the lower layer such as path planning completion and action execution success, forming a closed loop and driving subsequent tasks to continue execution. If an exception occurs during execution, an emergency process is triggered through the exception reporting node.

[0074] Emergency procedures are managed centrally by the emergency task module. When an anomaly occurs, the system records the current and last transition nodes, locates the anomaly, and initiates the emergency plan matching phase. The system queries the pre-set emergency plan database to determine if a matching response strategy exists. If a match is found, an emergency task is generated, involving operations such as replanning the path and adjusting the robotic arm's posture. This emergency task is then inserted into a waiting queue, typically given high priority for immediate processing, while other non-urgent tasks are suspended. If a match fails, the task is aborted, attempting manual intervention or system self-recovery. Subsequently, the system assesses the success of troubleshooting: if successful, normal procedures resume; if unsuccessful, manual intervention is triggered, requiring operators to manually intervene to ensure transformer testing safety.

[0075] The task priority and exception handling process of the task scheduling layer mentioned above realizes the fine-grained management of meta-tasks and dynamic response to exceptions. It not only ensures the sequential execution of regular tasks, but also allows for the rapid insertion of emergency tasks in case of emergencies, effectively improving the robustness and adaptability of the transformer detection system.

[0076] In this implementation, by introducing task dependencies, priority sorting, and dynamic emergency response mechanisms, the task scheduling layer can reasonably arrange the execution order of meta-tasks according to the constraints of the actual testing process, avoiding waiting or conflicts caused by improper task order. At the same time, when abnormal events such as instrument failure or power fluctuations occur during the testing process, the system can quickly call the preset emergency plan tasks, dynamically adjust the execution order of the remaining tasks, and prioritize operations such as abnormal recovery or safe shutdown, thereby greatly improving the production line's autonomous ability to cope with emergencies and the stability of continuous operation, and reducing manual intervention and downtime.

[0077] Further, the step of obtaining the transformer corresponding to the meta-task, and then performing instruction semantic parsing and key feature extraction on the meta-task based on the transformer and the robotic arm corresponding to the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, includes: Obtain the 3D digital model of the transformer corresponding to this meta-task, and the 3D digital model of the robotic arm corresponding to this meta-task. The instruction semantics of the meta-task are parsed to obtain the core actions, detection objects and task constraints corresponding to the meta-task. Key features are extracted from the meta-task to obtain the target position features, end effector attitude features, and motion parameter features corresponding to the meta-task. Based on the three-dimensional digital model of the transformer, the three-dimensional digital model of the robotic arm, the core motion, the detection object, the task constraints, the target position features, the end effector posture features, and the motion parameter features, model matching and coordinate transformation are performed to calculate the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task.

[0078] It should be understood that the Path Planning Layer (PPN) plays a crucial role in the entire layered architecture, bridging the gap between the layers. It receives abstract task instructions from the Task Scheduling Layer (TPN) and transforms them into concrete, parameterized motion trajectories that the Action Execution Layer (APN) can directly execute. In complex, confined, and obstacle-filled detection environments, the PPN generates a safe and efficient motion path for the robotic arm.

[0079] In one specific embodiment, the core functionality of the PPN is decomposed into three closely linked steps. The first step is instruction parsing and target extraction. The PPN receives instructions from the TPN and extracts key information and speed accuracy requirements. To transform the abstract task into precise spatial coordinates, the PPN accesses the three-dimensional digital models of the transformer and the robotic arm. Through model matching and coordinate transformation, it calculates the precise three-dimensional coordinates (x, y, z) and attitude (roll, pitch, yaw) of the target point in the robotic arm's base coordinate system.

[0080] Taking four typical tasks in transformer factory testing—DC resistance testing, airtightness testing, grounding current testing, and tap changer testing—as examples, the specific processing procedure for PPN is as follows: The first step is instruction semantic parsing. PPN first performs semantic parsing on the meta-task instructions to extract three key pieces of information: core action, detection object, and task constraints.

[0081] For the DC resistance test task, the core action is extracted as follows: move to the target position and clamp the test probe; the test object is: the low-voltage winding terminal of phase B; specifically described as: a copper columnar structure on the low-voltage side of the transformer, with a diameter of 12mm and a height of 30mm; the task constraints are: the test probe must make reliable contact with the conductive surface of the terminal, the contact pressure is 5±1N, and the contact point must be at the center of the top of the terminal to avoid the edge oxide layer affecting the test accuracy.

[0082] For the airtightness test task, the core action is extracted as follows: move to the flange face sealing groove and fit the test probe; the test object is: the annular sealing groove on the flange face of the oil tank, with parameters of "10mm wide, 5mm deep, located on the inner side of the flange edge, and 3000mm in circumference"; the task constraints are: the probe needs to move along the center trajectory of the sealing groove, with a fitting pressure of 20±2N and a moving speed of 30mm / s, to ensure that there is no air leakage in the sealing test.

[0083] For the grounding current test task, the core action is extracted as follows: move to the grounding terminal and fit the clamp meter; the test object is: iron core grounding terminal, located on the bottom side of the oil tank, cylindrical, with a diameter of 10mm, an exposed length of 50mm, and a horizontal axis; the task constraints are: the clamp meter jaws must completely cover the terminal, with a covering depth of ≥30mm, and the jaw plane must be perpendicular to the terminal axis.

[0084] For the tap changer test element task, the core actions are extracted as follows: move to the operating lever, grab, and rotate to switch gears; the test object is: tap changer operating lever, cylindrical, 25mm in diameter, with a hexagonal operating head on top, and 1800mm above the ground; the task constraints are: the grab position is 10mm below the operating head to avoid slippage, the rotation angle must be precise, each gear corresponds to a 30° rotation, and from gear 1 to gear 3, a 60° clockwise rotation is required.

[0085] Step 2: Key Feature Extraction. Key features are extracted from the meta-task to obtain the target position features, end effector attitude features, and motion parameter features corresponding to the meta-task.

[0086] For the DC resistance testing task, the extracted target position features include: the center of the terminal tip (center of the conductive surface, test contact point) has coordinates (Xt=1250mm, Yt=800mm, Zt=1500mm) in the transformer design coordinate system; the terminal axis direction is perpendicular to the winding plane (positive Z-axis direction), i.e., the terminal is vertically upward. The end effector attitude features include: the tool coordinate system Z-axis coincides with the terminal axis, and the tool coordinate system X-axis is radially aligned with the terminal. Motion parameter features include: approach speed 50mm / s; positioning accuracy ±0.02mm.

[0087] For the airtightness test element task, the extracted target location features include: the center trajectory of the sealing groove is a ring, and the center of the ring is located in the transformer coordinate system (Xt=1000mm, Yt=1000mm, Zt=2000mm); the radius of the ring is R=500mm, and the plane on which it is located is parallel to the top surface of the transformer (Zt=2000mm plane).

[0088] The end effector's attitude characteristics include: the tool's Z-axis is perpendicular to the flange face (i.e., parallel to the transformer coordinate system's Z-axis, pointing downwards towards the flange face); the tool's X-axis is along the tangent direction of the sealing groove (ensuring the probe's movement direction is consistent with the trajectory). Motion parameter characteristics include: contact pressure 20N; trajectory accuracy ±0.1mm.

[0089] For the grounding current test element task, the extracted target position features include: the midpoint of the exposed part of the terminal, the center position of the clamp meter, and the coordinates in the transformer coordinate system (Xt=800mm, Yt=1200mm, Zt=500mm); the terminal axis direction is along the negative Y-axis of the transformer coordinate system (horizontally to the left). The end effector attitude features include: the tool's X-axis coincides with the terminal axis (negative Y-axis direction); the tool's Y-axis is perpendicular to the terminal axis (horizontal direction); the tool's Z-axis is vertically upward (ensuring the jaw opening direction is horizontal to avoid interference with the ground). Motion parameter features include: insertion speed 20mm / s, jaw opening degree ≥15mm.

[0090] For the tap changer test element task, the extracted target position features include: the center of the operating lever gripping (10mm from the top operating head), in the transformer coordinate system (Xt=1500mm, Yt=900mm, Zt=1800mm); the operating lever axis direction is vertical (positive Z-axis direction). The end effector attitude features include: the tool Z-axis coincides with the operating lever axis (vertical direction); the tool X-axis points to the right (avoiding the left oil pipe, the clamp opening faces right). Motion parameter features include: gripping torque 3 N·m; rotational accuracy ±1°.

[0091] The third step involves PPN accessing the 3D digital models of the transformer and the robotic arm, combining the actual installation parameters of the transformer, transforming the target point from the transformer design coordinate system to the robotic arm base coordinate system, and calculating the final attitude angles (roll, pitch, yaw).

[0092] For the DC resistance test element task, the calculated translation parameters are: X0=500mm, Y0=300mm, Z0=1000mm. The rotation parameters are: 0° rotation around the Z-axis of the base coordinate system. The attitude is transformed into: roll=0°, pitch=0°, yaw=0°.

[0093] For the airtightness test element task, the calculated translation parameters are: X0=500mm, Y0=500mm, Z0=1500mm. The rotation parameters are: rotation around the Z-axis by 90° (the transformer axis is parallel to the X-axis of the base coordinate system). The attitude is transformed into: roll=0°, pitch=90°, yaw=θ.

[0094] For the ground current test element task, the calculated translation parameters are: X0=1000mm, Y0=800mm, Z0=300mm. The rotation parameters are: no rotation (coordinate system fully aligned). The attitude is transformed into: roll=0°, pitch=0°, yaw=90°.

[0095] For the tap changer test element task, the calculated translation parameters are: X0=500mm, Y0=500mm, Z0=0mm. The rotation parameters are: 0° rotation around the X-axis (horizontal placement). The attitude is converted to: roll=0°, pitch=0°, yaw=0°.

[0096] Through the above three steps, PPN transforms the meta-task instructions of the four detection items into precise task coordinate parameters (x, y, z) and task pose parameters (roll, pitch, yaw), providing complete and accurate input for the Action Execution Layer (APN) to generate specific motion trajectories.

[0097] In this implementation, three-dimensional digital models and semantic parsing technology are used to transform abstract meta-task instructions into specific geometric parameters. This enables the path planning layer to accurately understand the task intent and automatically adapt to the size and structural differences of different transformer models and robotic arms. No manual teaching or repeated debugging is required, which significantly improves the versatility and accuracy of path planning and reduces the deployment cycle and technical threshold.

[0098] Further, generating the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and the task posture parameters includes: The path start point and path end point are obtained based on the task coordinate parameters and the task attitude parameters; Based on the global search algorithm and the fast expanding random tree algorithm, the geometric path from the starting point of the path to the ending point of the path is obtained; The dynamic model and kinematic constraints of the robotic arm are obtained, and then the geometric path is interpolated and optimized based on the dynamic model and kinematic constraints to generate a motion trajectory.

[0099] For collision-free path planning, after obtaining the starting and ending positions of the robotic arm, the PPN searches the configuration space to find a continuous path connecting the starting and ending points. The configuration space is a high-dimensional space with dimensions equal to the robotic arm's degrees of freedom. In this space, each point represents a possible pose of the robotic arm. The PPN searches in this high-dimensional space while avoiding obstacle areas formed by transformer structures and its own links. The path planning employs an A* global search combined with a RRT (Rapidly Expanded Random Tree) algorithm to efficiently find a feasible path in complex environments.

[0100] For trajectory generation and optimization, the path planning algorithm outputs only a geometric path composed of discrete points; the robotic arm cannot instantly jump from one point to another. PPN transforms this geometric path into a smooth motion trajectory that varies over time. This process is called trajectory generation parameterization. Based on the robotic arm's dynamic model and kinematic constraints, PPN interpolates and optimizes the path, generating a time-optimal or energy-optimal trajectory that satisfies all constraints. This trajectory specifies in detail the position, velocity, and acceleration of the robotic arm's end effector at each moment, as well as the angles, angular velocities, and angular accelerations of each joint. This detailed trajectory instruction is ultimately passed to APN for execution.

[0101] In one specific embodiment, in the control method for a transformer detection robotic arm based on a hierarchical architecture, the Path Planning Layer (PPN) undertakes the core responsibility of transforming the meta-task into a specific motion trajectory. To achieve efficient, safe, and smooth path search and optimization, the PPN employs a hybrid algorithm combining a global search algorithm and a Rapidly Expanding Random Tree (RRT), and achieves their collaborative operation through a timing control mechanism. This process encompasses the entire chain from cost function design, path search, and global path generation to local optimization, physical verification, and task transfer, which can be elaborated as follows.

[0102] The starting point for path search is the design of the cost function, which determines the evaluation criteria for path quality during the search process. To this end, PPN designs two parts: the actual cost function and the heuristic cost function. The actual cost function... Taking into account both the cost of joint movement and the safety cost of obstacles, the formula is as follows: Weighted by the difference in joint angles. For obstacle safety weights, For the first Path points, for The previous path point, for The minimum distance between a point and an obstacle. For a multi-axis robotic arm, the number of its joint axes is equal to the total number of path points in the actual cost function. In this embodiment, a six-axis robotic arm is used as an example to calculate the actual cost function; therefore, the total number of path points in the actual cost function is 6.

[0103] A heuristic cost function is designed using weighted Euclidean distance, taking into account both joint angle differences and end-effector position differences: In the formula As the key angle difference weight, Weighted by the difference in terminal positions. for The next path point, , , for Point rectangular coordinate components, , , for The Cartesian coordinate components of the next path point. For a multi-axis robotic arm, the number of joint axes it has is equal to the total number of path points in the heuristic cost function. In this embodiment, a six-axis robotic arm is used as an example to calculate the heuristic cost function; therefore, the total number of path points in the heuristic cost function is 6.

[0104] During the search process, PPN maps the steps of the global search algorithm to places and transitions in a Petri net (PPN), forming a discrete event-driven search mechanism. Initially, the starting place... Injection of the carryover point Tokens, OpenList library Add to the starting point, close the list library. Empty, obstacle library Load the configuration space coordinates of the obstacles. Then trigger the optimal node selection transition. Select the total cost from the open list. The smallest node is selected and moved to the off list. Then, a neighbor node transition is triggered. For the current node, generate neighboring nodes whose six joints have changed by ±0.5°, and perform transition detection through collision detection. Verify its feasibility. Finally, trigger the cost update transition. Calculate the actual cost for feasible neighbor nodes. And update the open list, in the formula It is the current point. It is the first Neighboring nodes of each path point It is the first There are several path points. Through this series of iterative transitions, the PPN eventually outputs a global path that meets the basic requirements of security and path continuity.

[0105] The global path has clear characteristic parameters: the total joint angle change does not exceed 320°, the end-effector movement trajectory is continuous, the minimum distance between the robotic arm and obstacles corresponding to all nodes on the path is not less than 60mm, and the time taken for the entire search process is controlled within 200ms to ensure real-time requirements.

[0106] However, the paths output by the global search algorithm consist of discrete nodes, which is insufficient for direct use in the motion control of the robotic arm. Therefore, PPN introduces the Restricted Tracking Response (RRT) algorithm for local optimization and trajectory smoothing. The adaptation of RRT is mainly reflected in three aspects: first, sampling area limitation, sampling is only performed within ±5° of the joint angles around path A, with a step size of 0.2°, thus focusing on the feasible space within the global path neighborhood; second, cost function enhancement, by adding joint angular velocity constraints. and acceleration constraints First, avoid motion overshoot; second, adjust the rewire strategy to prioritize parent nodes with small joint angle changes to ensure path smoothness.

[0107] In the timing control optimization of PPN, the RRT extension process is also modeled using Petri nets. The optimization of the startup library... Receive path A token to trigger RRT initialization transition. Generate an initial tree. Then trigger a random sampling transition. Generate random nodes within the constrained region. The feasibility was verified through collision detection. Then, a search for the nearest node to determine the transition was triggered. Find the nearest node in the current tree. Then trigger node expansion transition. exist Towards Generate new nodes in the direction If the node satisfies the kinematic constraints, it is added to the tree. This then triggers the Rewire transition. ,right Reconnect to surrounding nodes if... Reach the target point If the cost is lower and the path is smoother, then update the parent node. When the joint angle difference between the generated path and the target point is less than 0.1°, trigger the optimization to complete the transition. Output optimized path .

[0108] The optimized path has more stringent kinematic characteristics: the rate of change of joint angle does not exceed 3° / step, the end effector trajectory is continuous and has a circular transition, and the curve is smooth without convexity; the angular velocity of all joints does not exceed 60° / s, and the acceleration does not exceed 60° / s², ensuring that the movement of the robotic arm in physical space is stable and controllable.

[0109] To further verify the feasibility of the path, PPN also needs to undergo physical space collision verification. The optimized path will then be used. Each joint angle combination is converted into physical coordinates of the end effector and links through forward kinematics. The minimum distance between all links and the transformer housing, testing instruments, and transmission track is checked to ensure it meets the safety requirement of ≥60mm, and the trajectory of the end effector from the starting point to the ending point is free from interference. The verification process can simulate the optimized path motion of the robotic arm using simulation platforms such as Unity, visually confirming the feasibility and safety of the path.

[0110] Ultimately, the PPN and the Action Execution Layer (APN) achieve task coordination through timing. After the PPN completes path planning, it triggers path availability transitions. Generate an action preparation library containing an information token and transmit it to the APN. .

[0111] The token contains a sequence of waypoints. The system includes motion parameters (such as joint velocity and acceleration) and safety constraint information (such as the minimum distance to obstacles corresponding to the path segment). After receiving the token, the APN triggers the action to execute the transition. The robotic arm is controlled to complete a collision-free motion task following an optimized path. Thus, the entire process from task analysis and path planning to motion execution is achieved in a closed loop.

[0112] In this implementation, by integrating a global search algorithm with a fast expanding random tree algorithm, the method can quickly search for collision-free geometric paths while ensuring global optimality. This solves the problem that single algorithms are prone to getting trapped in local optima or having low search efficiency in complex environments. Furthermore, by combining the dynamic model and kinematic constraints of the robotic arm, the geometric path is interpolated and optimized to ensure that the generated motion trajectory meets physical constraints such as joint velocity and acceleration. This ensures that the robotic arm can perform tasks smoothly and stably, extends equipment life, and improves operational safety.

[0113] Furthermore, obtaining the geometric path from the starting point to the ending point of the path based on the global search algorithm and the fast expanding random tree algorithm includes: Obtain the comprehensive joint motion cost and obstacle safety cost of the corresponding robotic arm, and then construct the actual cost function based on the comprehensive joint motion cost and the obstacle safety cost; Obtain the joint angle difference and end-effector position difference corresponding to the robotic arm, and then construct a heuristic cost function based on the joint angle difference and the end-effector position difference; A total cost function is constructed based on the actual cost function and the heuristic cost function, and joint angular velocity constraints and acceleration constraints are added to the total cost function. Based on the total cost function, the path start point, and the path end point, a path iterative search is performed in the preset configuration space of the robotic arm to obtain the geometric path from the path start point to the path end point.

[0114] Furthermore, the step of instructing the action execution layer to control the movement of the corresponding robotic arm based on the motion trajectory, so that the robotic arm completes the corresponding meta-task, includes: Based on the motion trajectory, the desired angle, desired angular velocity, and desired angular acceleration of each joint of the robotic arm are obtained; The actual angle, actual angular velocity, and actual angular acceleration of each joint of the robotic arm are detected, and then the angle error between the desired angle and the actual angle, the angular velocity error between the desired angular velocity and the actual angular velocity, and the angular acceleration error between the desired angular acceleration and the actual angular acceleration are calculated respectively. Based on the angle error, the angular velocity error, and the angular acceleration error, a joint control signal is generated using a PID control algorithm. Then, the movement of the corresponding robotic arm is controlled according to the joint control signal so that the robotic arm can complete the corresponding meta-task.

[0115] It's important to note that the Action Execution Layer (APN) is the final execution stage in the entire hierarchical control architecture. It directly controls every joint of the robotic arm, translating the abstract motion trajectory from the Procedural Processing Layer (PPN) into concrete physical movements. The core task of the APN is to ensure that the robotic arm can track the desired motion trajectory in real time with extremely high precision and stability, while simultaneously collecting feedback information from various sensors to provide data support for upper-level control. Its performance directly determines the success or failure of the entire detection task.

[0116] The core functionality of the APN is divided into three levels. First, it handles trajectory command reception and parsing. The APN receives detailed motion trajectory commands from the PPN. This command is a time series containing the desired angle, angular velocity, and angular acceleration of each joint of the robotic arm in each control cycle. The APN accurately parses these commands and uses them as reference inputs for its own control loop. Second, it performs real-time joint motion control. The APN runs on a motion control card to ensure control accuracy and stability, employing a high-speed closed-loop feedback control mechanism. Within each control cycle, the APN reads the actual angle of the joint in real time via encoders mounted on each joint motor. Then, it compares the actual angle with the desired angle from the trajectory command, calculating the angle error. Based on this error, the APN's internal PID controller calculates a control signal, which is sent to the joint motor via a servo driver, driving the motor to rotate and reduce the error. This process is repeated continuously at a very high frequency, ensuring that the robotic arm's actual movement closely follows the desired trajectory.

[0117] Finally, there's sensor data acquisition and processing. While performing motion, the APN is also responsible for collecting and processing data from various sensors, including proprioceptive data such as joint torque, current, and temperature, used to monitor the robot arm's own health. External sensing data includes data collected by force / torque sensors and vision sensors mounted on the end effector. This data is crucial for achieving more advanced control functions. For example, force data can be used for force control, enabling the robot arm to control the magnitude of contact force when in contact with transformer components, preventing damage to the equipment. Vision data can be used for visual servoing, adjusting the robot arm's position in real time to align with the target.

[0118] The APN performs preliminary processing and filtering on this raw data, and then packages key status information and abnormal events (such as collision detection, motor overload, encoder failure, etc.) and feeds them back to the upper-level PPN and TPN in real time through the communication interface, forming a complete closed-loop control system.

[0119] The APN technology implementation system comprises two core modules: an embedded motion controller and a servo drive system. These two modules establish closed-loop communication via an industrial real-time bus. The embedded motion controller parses the APN's motion token commands, executes real-time control algorithms, and outputs low-power control signals. The servo drive system receives these control signals, amplifies them, and drives the servo motors to perform actions. It also feeds back real-time position and speed information to the embedded motion controller via an encoder, forming a closed-loop control system of command generation, execution, feedback, and correction. Ultimately, this achieves the robotic arm's motion sequence defined by the APN.

[0120] The embedded motion controller, employing a heterogeneous multi-core hardware architecture and a layered functional module design, possesses real-time control, parallel processing of non-real-time tasks, industrial-grade anti-interference capabilities, and data security. The servo drive system is responsible for converting the low-power control signals from the embedded motion controller into high-power drive signals, driving the servo motors to perform actions, and providing real-time feedback on the motion status.

[0121] In the layered architecture-based control method for transformer inspection robotic arms, the Action Execution Layer (APN) is the final execution stage of the entire control system. Its core technology relies on the collaborative work of an embedded motion controller and a servo drive system. The two form a closed-loop control circuit via an industrial real-time bus (such as EtherCAT), ensuring the robotic arm can track motion trajectory commands from the Path Planning Layer (PPN) with high precision and real-time performance. The embedded motion controller, as the core control unit, adopts a heterogeneous multi-core hardware architecture and integrates multiple functional modules, each responsible for tasks such as real-time control, non-real-time task processing, communication interaction, and fault diagnosis.

[0122] In the real-time control module, the controller first parses the path token from the PPN, which contains the target quasi-pose and motion velocity parameters of the robotic arm's end effector. Inverse kinematics is then performed using the DH parameter method to convert the end effector pose into target angles for the six joints, with the solution error controlled within 0.001°. Subsequently, an S-curve interpolation algorithm is used to smoothly interpolate the joint angles, limiting the rate of change of joint angular velocity to no more than 5° / ms to eliminate jitter in the robotic arm's movements. The control cycle is 200μs. Within each cycle, the system completes closed-loop control of "command output – feedback acquisition – deviation correction," using a PID algorithm to correct joint angle deviations in real time, ensuring the accuracy of trajectory tracking.

[0123] The non-real-time task processing module is responsible for managing background tasks that run concurrently with motion control. This includes storing APN action logs in a standard format, with log content including timestamps, action numbers, and deviation values; receiving task priority instructions from the Task Scheduling Layer (TPN) via the Profinet interface and feeding back the current motion status of the robotic arm to the PPN; and classifying, storing, and analyzing fault signals fed back by the servo drive system to generate fault reports for subsequent maintenance and optimization.

[0124] The communication module handles data exchange with the servo drive system and the detection platform. Communication with the servo drive system is via EtherCAT, using the CoE protocol to transmit joint target angle and speed commands, as well as encoder feedback data, with a packet loss rate of less than 0.1%. Communication with the detection platform uses the HTTPS protocol for data transmission, employing the national cryptographic SM2 algorithm for authentication and the SM4 algorithm to encrypt APN motion parameters, ensuring data transmission security. The fault diagnosis module monitors the controller hardware status and external feedback signals in real time, promptly detecting anomalies and triggering corresponding processing mechanisms.

[0125] The servo drive system, acting as the execution unit of the embedded motion controller, receives and amplifies low-power control signals to drive the servo motor. Its core control mechanism is a three-loop control structure, including an inner current loop, a middle speed loop, and an outer position loop. The inner current loop samples the motor stator current at a frequency of 10kHz, using a PI controller to control the difference between the actual current and the target current within ±0.5A, with a bandwidth of no less than 1kHz, effectively suppressing torque fluctuations. The middle speed loop samples the motor speed at a frequency of 1kHz, using a PI controller to control the speed difference within ±5rpm, with a bandwidth of no less than 200Hz, avoiding speed overshoot. The outer position loop samples the motor rotor position at a frequency of 100Hz, using a PID controller to control the position difference within ±0.001°, corresponding to a robotic arm end-effector positioning accuracy of ±0.01mm.

[0126] In addition, the servo drive system features multiple adaptive compensation functions to enhance control performance. Load disturbance compensation adjusts the speed loop output in real time based on torque changes fed back from the current loop, keeping speed fluctuations within 5 rpm when the robotic arm increases its gripping load from 10 kg to 20 kg. Friction compensation pre-stores the motor's friction characteristic curve, outputting an additional 2 N·m of torque at low speeds (<50 rpm) to eliminate creeping. Temperature compensation monitors the motor winding temperature, automatically adjusting the current loop PI parameters when the temperature rises from 25°C to 60°C to ensure stable torque output.

[0127] In summary, the embedded motion controller and servo drive system, through modular design and high real-time collaboration, jointly achieved high-precision control of the robotic arm's movement at the APN layer, providing a solid technical guarantee for the successful completion of the transformer testing task.

[0128] The aforementioned layered architecture-based control method for a transformer detection robotic arm achieves task decoupling and modular design. It decomposes the complex control problem into independent layers, including task scheduling, path planning, and motion execution, with each layer designed, developed, and tested independently. The task scheduling layer focuses on optimizing the detection sequence, without concern for the underlying motor control details. The motion execution layer focuses on achieving high-precision trajectory tracking, without concern for the global detection target. This modular design significantly reduces system complexity and improves code reusability and maintainability. When the system is upgraded or a functional module is replaced, only the corresponding layer needs modification, reducing development and maintenance costs.

[0129] Secondly, the layered architecture significantly enhances the system's robustness and scalability. In this structure, each layer possesses a degree of autonomy and fault tolerance. When the action execution layer detects a slight trajectory deviation, it quickly corrects it using local control algorithms without reporting to the task scheduling layer, thus ensuring the system's real-time performance and stability. When encountering severe, unmanageable faults, the error message is propagated upwards, where a higher-level planner makes a decision, replanning the path or adjusting the task sequence. This hierarchical approach to exception handling allows the system to more gracefully cope with various unforeseen circumstances. Furthermore, the layered architecture facilitates system expansion. Adding new detection functions simply requires adding the corresponding task logic to the task scheduling layer and providing corresponding motion support in the path planning and action execution layers, without needing to reconstruct the entire control system.

[0130] Please refer to Figure 5 The second aspect of this invention provides a transformer detection robotic arm control system based on a hierarchical architecture, used to control a cluster of robotic arms, including a task scheduling layer 100, a path planning layer 200, and an action execution layer 300; wherein: The task scheduling layer is used to obtain transformer detection instructions, obtain macro tasks based on the transformer detection instructions, and then decompose the macro tasks into several meta tasks according to the preset transformer detection procedures, preset standard operating procedures and the model and status of each robotic arm in the robotic arm cluster. Then, a robotic arm is assigned to each meta task, and the several meta tasks are sent to the path planning layer. The path planning layer is used for any of the meta-tasks: obtaining the transformer corresponding to the meta-task, and based on the transformer and the robotic arm corresponding to the meta-task, performing instruction semantic parsing and key feature extraction on the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, and then generating the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and the task posture parameters, and sending the motion trajectory to the action execution layer. The motion execution layer is used to control the movement of the corresponding robotic arm based on the motion trajectory using a PID control algorithm, so that the robotic arm can complete the corresponding meta-task.

[0131] It should be noted that the system's software architecture corresponds to its hardware architecture, also employing a layered design to achieve modular functionality and clear interface definitions. The system's perception layer includes sensor drivers, data preprocessing, and multimodal information fusion modules, used to collect and fuse data from different sensors to construct a unified environmental model.

[0132] The system processing layer includes the TPN module, PPN module, and task decoupling and program generation engine, which are responsible for task planning, path planning, and instruction generation, and are the core of the system's decision-making.

[0133] The system communication layer includes the real-time data bus EtherCAT, the network protocol TCP / IP, and ROS, which are used to achieve efficient and reliable data interaction and command transmission between different layers.

[0134] The perception layer software manages and processes all sensor data, including drivers for various sensors, data filtering and preprocessing algorithms, and a multimodal information fusion module. The system fuses point cloud data generated by LiDAR with color images acquired by cameras to create a 3D environment model with both precise geometric and rich texture information. This model serves as the unified environment model for path planning in the upper-layer PPN. The environment model information includes three categories: geometric information, obstacle information, and feasible region information. Geometric information, acquired from LiDAR point clouds, includes the 3D coordinates and boundaries of the workstation, providing the robotic arm configuration space (C-space) for the PPT. Obstacle information, obtained through image recognition from the camera, provides differentiated constraints for the PPN, including obstacle type, safety threshold, and dynamic shape. Feasible region information, generated by path search and optimization algorithms, provides motion path points and movable ranges for the PPT.

[0135] The processing layer is the core of the software architecture, containing the main logic of TPN and PPN. The TPN module is responsible for task parsing, decomposition, and scheduling. The PPN module is responsible for path planning and trajectory generation. In addition, a task decoupling and program generation engine is introduced, which automatically generates the corresponding PPN planning program and APN control program based on the task sequence generated by TPN.

[0136] The communication layer ensures seamless collaboration between other layers. Between the APN and the lower-level hardware, the highly real-time industrial fieldbus EtherCAT is used to guarantee low-latency transmission of control commands and sensor data. Between the host computer and the lower-level hardware, and between the TPN, PPN, and APN software modules, the standard network protocol TCP / IP is employed. In the robotic arm ROS system, each module is designed as an independent node, communicating loosely through a publishing mechanism, greatly improving the flexibility and scalability of software development.

[0137] Further, the task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet; wherein, sending a plurality of the meta-tasks to the path planning layer includes: A task activation token is generated, and a meta-task repository is established for each meta-task based on the task activation token. Then, each meta-task repository is passed to the path planning initiation repository of the path planning Petri subnet through the corresponding substitution transition, triggering the calculation logic of the path planning Petri subnet. Sending the motion trajectory to the action execution layer includes: A path availability status token is generated, and then the path availability status token is returned to the task scheduling Petri subnet through the port library. The motion trajectory is also transmitted to the action execution Petri subnet through the port library, so that the task scheduling Petri subnet issues an execution token corresponding to the motion trajectory to the action execution Petri subnet based on the path availability status token.

[0138] Furthermore, the task scheduling layer is used to obtain transformer detection instructions, obtain macro-level tasks based on the transformer detection instructions, and then decompose the macro-level tasks into several meta-tasks according to the preset transformer detection procedures, preset standard operating procedures, and the model and status of each robotic arm in the robotic arm cluster. Then, a robotic arm is assigned to each meta-task, and the several meta-tasks are sent to the path planning layer, including: Obtain the logical dependencies and execution priorities between each of the meta-tasks, and then determine the task execution order of several meta-tasks based on the logical dependencies and execution priorities; In response to any abnormal execution result of the meta-task, the emergency response plan task corresponding to the abnormal execution result is obtained, and then the task execution order is updated based on the emergency response plan task; A robotic arm is assigned to each of the meta-tasks according to the task execution order, and several of the meta-tasks are sent to the path planning layer.

[0139] The present invention provides a control method and system for a transformer detection robotic arm based on a hierarchical architecture, which has at least the following advantages compared to the prior art: First, this invention constructs a three-layer closed-loop architecture consisting of a Task Scheduling Network (TPN), a Path Planning Network (PPN), and an Action Execution Network (APN), to achieve precise conversion and hierarchical coordination from macroscopic detection commands to the physical actions of the robotic arm, thus solving the problem that existing single-machine level control cannot meet the cycle management requirements of multi-robotic arm systems.

[0140] Secondly, this invention leverages Petri nets' transition and place mechanisms to achieve modular decomposition of the layered architecture, event-driven interaction between layers, and timing control. Simultaneously, it uses verification methods to mitigate deadlock and collision risks, ensuring reliable system operation. The invention includes methods for modular decomposition of the layered architecture based on Petri nets, layered interaction, timing control, and deadlock prevention and obstacle avoidance safety verification techniques.

[0141] Finally, this invention integrates embedded motion controllers, servo drive systems, and dedicated path optimization and sensor fusion algorithms to support high-precision and high-stability detection of robotic arms. It uses A global search and RRT fast expansion algorithms to optimize paths and combines multimodal sensor fusion to achieve end-effector positioning accuracy, thus solving the problems of existing hardware redundancy and insufficient motion accuracy.

[0142] Ultimately, this invention, through the aforementioned hierarchical architecture-based control method for transformer inspection robotic arms, significantly improves hardware resource utilization, reduces cost redundancy, and breaks through the existing "one arm per station" independent configuration mode. Based on a directed graph structure and token flow mechanism, it constructs a centralized control system for multiple robotic arms, achieving flexible collaboration across workstations. The prototype experimental line requires only 50% of the original hardware configuration to achieve the same testing capabilities, solving the problems of large differences in robotic arm utilization rates across different workstations, severe hardware idleness, and high maintenance costs.

[0143] Meanwhile, the timing control flexibility and fault tolerance have been significantly optimized by establishing a hierarchical timing coordination mechanism, which can dynamically adapt to fluctuations in detection duration and adjust the control logic according to the busy / idle status of the workstation. In the face of emergencies such as instrument failure and power fluctuations, breakpoint resume can be achieved without manual intervention, solving the problems of rigid control timing and asynchronous states in the existing system.

[0144] Furthermore, by replacing traditional physical isolation with a multimodal perception fusion system, and combining global search and RRT (Rapidly Extended Path Planning) algorithms, real-time collision prediction and interference-free obstacle avoidance of the robotic arm's movement path are achieved. The average obstacle avoidance time is far less than the 10 minutes of existing technologies, and the minimum distance between the robotic arm and obstacles on the path is ≥60mm, significantly reducing the risk of collision.

[0145] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0146] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; however, any combination of these technical features that does not contradict each other should be considered within the scope of this specification.

[0147] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the concept of this application, and these improvements and substitutions should also be considered within the scope of protection of this invention. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A control method for a transformer detection robotic arm based on a hierarchical architecture, used to control a cluster of robotic arms, characterized in that, This method is based on a pre-defined layered architecture, which includes a task scheduling layer, a path planning layer, and an action execution layer. The method includes the following steps: The task scheduling layer responds to the transformer detection command, obtains the macro task based on the transformer detection command, and then decomposes the macro task into several meta tasks according to the preset transformer detection procedure, the preset standard operating procedure, and the model and status of each robot arm in the robot arm cluster. Then, a robot arm is assigned to each meta task, and several meta tasks are sent to the path planning layer. For any given meta-task, the path planning layer obtains the transformer corresponding to the meta-task, performs instruction semantic parsing and key feature extraction on the meta-task based on the transformer and robotic arm corresponding to the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, and then generates the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and task posture parameters, and sends the motion trajectory to the action execution layer. The motion execution layer controls the movement of the corresponding robotic arm based on the motion trajectory, so that the robotic arm completes the corresponding meta-task.

2. The control method for a transformer detection robotic arm based on a hierarchical architecture according to claim 1, characterized in that, The task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet; wherein, sending a plurality of the meta-tasks to the path planning layer includes: A task activation token is generated, and a meta-task repository is established for each meta-task based on the task activation token. Then, each meta-task repository is passed to the path planning initiation repository of the path planning Petri subnet through the corresponding substitution transition, triggering the calculation logic of the path planning Petri subnet. Sending the motion trajectory to the action execution layer includes: A path availability status token is generated, and then the path availability status token is returned to the task scheduling Petri subnet through the port library. The motion trajectory is also transmitted to the action execution Petri subnet through the port library, so that the task scheduling Petri subnet issues an execution token corresponding to the motion trajectory to the action execution Petri subnet based on the path availability status token.

3. The control method for a transformer detection robotic arm based on a hierarchical architecture according to claim 1, characterized in that, The task scheduling layer responds to the transformer detection command, obtains the macro-task based on the transformer detection command, and then decomposes the macro-task into several meta-tasks according to the preset transformer detection procedure, the preset standard operating procedure, and the model and status of each robotic arm in the robotic arm cluster. Subsequently, a robotic arm is assigned to each meta-task, and the several meta-tasks are sent to the path planning layer, including: Obtain the logical dependencies and execution priorities between each of the meta-tasks, and then determine the task execution order of several meta-tasks based on the logical dependencies and execution priorities; In response to any abnormal execution result of the meta-task, the emergency response plan task corresponding to the abnormal execution result is obtained, and then the task execution order is updated based on the emergency response plan task; A robotic arm is assigned to each of the meta-tasks according to the task execution order, and several of the meta-tasks are sent to the path planning layer.

4. The control method for a transformer detection robotic arm based on a hierarchical architecture according to claim 1, characterized in that, The process involves obtaining the transformer corresponding to the meta-task, and based on the transformer and the robotic arm, performing instruction semantic parsing and key feature extraction on the meta-task to obtain the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, including: Obtain the 3D digital model of the transformer corresponding to the meta-task, and the 3D digital model of the robotic arm corresponding to the meta-task. The instruction semantics of the meta-task are parsed to obtain the core actions, detection objects, and task constraints corresponding to the meta-task. Key features are extracted from the meta-task to obtain the target position features, end effector attitude features, and motion parameter features corresponding to the meta-task. Based on the three-dimensional digital model of the transformer, the three-dimensional digital model of the robotic arm, the core motion, the detection object, the task constraints, the target position features, the end effector posture features, and the motion parameter features, model matching and coordinate transformation are performed to calculate the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task.

5. The control method for a transformer detection robotic arm based on a hierarchical architecture according to claim 1, characterized in that, The process of generating the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and the task posture parameters includes: The path start point and path end point are obtained based on the task coordinate parameters and the task attitude parameters; Based on the global search algorithm and the fast expanding random tree algorithm, the geometric path from the starting point of the path to the ending point of the path is obtained; The dynamic model and kinematic constraints of the robotic arm are obtained, and then the geometric path is interpolated and optimized based on the dynamic model and kinematic constraints to generate a motion trajectory.

6. The control method for a transformer detection robotic arm based on a hierarchical architecture according to claim 5, characterized in that, The process of obtaining the geometric path from the starting point to the ending point of the path based on the global search algorithm and the fast expanding random tree algorithm includes: Obtain the comprehensive joint motion cost and obstacle safety cost of the corresponding robotic arm, and then construct the actual cost function based on the comprehensive joint motion cost and the obstacle safety cost; Obtain the joint angle difference and end-effector position difference corresponding to the robotic arm, and then construct a heuristic cost function based on the joint angle difference and the end-effector position difference; A total cost function is constructed based on the actual cost function and the heuristic cost function, and joint angular velocity constraints and acceleration constraints are added to the total cost function. Based on the total cost function, the path start point, and the path end point, a path iterative search is performed in the preset configuration space of the robotic arm to obtain the geometric path from the path start point to the path end point.

7. The control method for a transformer detection robotic arm based on a hierarchical architecture according to claim 1, characterized in that, The step of instructing the action execution layer to control the movement of the corresponding robotic arm based on the motion trajectory, so that the robotic arm completes the corresponding meta-task, includes: Based on the motion trajectory, the desired angle, desired angular velocity, and desired angular acceleration of each joint of the robotic arm are obtained; The actual angle, actual angular velocity, and actual angular acceleration of each joint of the robotic arm are detected, and then the angle error between the desired angle and the actual angle, the angular velocity error between the desired angular velocity and the actual angular velocity, and the angular acceleration error between the desired angular acceleration and the actual angular acceleration are calculated respectively. Based on the angle error, the angular velocity error, and the angular acceleration error, a joint control signal is generated using a PID control algorithm. Then, the movement of the corresponding robotic arm is controlled according to the joint control signal so that the robotic arm can complete the corresponding meta-task.

8. A control system for a transformer detection robotic arm based on a hierarchical architecture, used to control a cluster of robotic arms, characterized in that, It includes a task scheduling layer, a path planning layer, and an action execution layer; among which: The task scheduling layer is used to obtain transformer detection instructions, obtain macro tasks based on the transformer detection instructions, and then decompose the macro tasks into several meta tasks according to the preset transformer detection procedures, preset standard operating procedures and the model and status of each robotic arm in the robotic arm cluster. Then, a robotic arm is assigned to each meta task, and the several meta tasks are sent to the path planning layer. The path planning layer is used for any of the meta-tasks: obtaining the transformer corresponding to the meta-task, and based on the transformer and the robotic arm corresponding to the meta-task, performing instruction semantic parsing and key feature extraction on the meta-task, thereby obtaining the task coordinate parameters and task posture parameters of the robotic arm corresponding to the meta-task, and then generating the motion trajectory of the robotic arm corresponding to the meta-task based on the task coordinate parameters and the task posture parameters, and sending the motion trajectory to the action execution layer. The motion execution layer is used to control the movement of the corresponding robotic arm based on the motion trajectory using a PID control algorithm, so that the robotic arm can complete the corresponding meta-task.

9. The control system for a transformer detection robotic arm based on a hierarchical architecture according to claim 8, characterized in that, The task scheduling layer includes a task scheduling Petri subnet, the path planning layer includes a path planning Petri subnet, and the action execution layer includes an action execution Petri subnet; wherein, sending a plurality of the meta-tasks to the path planning layer includes: A task activation token is generated, and a meta-task repository is established for each meta-task based on the task activation token. Then, each meta-task repository is passed to the path planning initiation repository of the path planning Petri subnet through the corresponding substitution transition, triggering the calculation logic of the path planning Petri subnet. Sending the motion trajectory to the action execution layer includes: A path availability status token is generated, and then the path availability status token is returned to the task scheduling Petri subnet through the port library. The motion trajectory is also transmitted to the action execution Petri subnet through the port library, so that the task scheduling Petri subnet issues an execution token corresponding to the motion trajectory to the action execution Petri subnet based on the path availability status token.

10. The control system for a transformer detection robotic arm based on a hierarchical architecture according to claim 8, characterized in that, The task scheduling layer is used to obtain transformer detection instructions, obtain macro-tasks based on the transformer detection instructions, and then decompose the macro-tasks into several meta-tasks according to preset transformer detection procedures, preset standard operating procedures, and the model and status of each robotic arm in the robotic arm cluster. Subsequently, a robotic arm is assigned to each meta-task, and the several meta-tasks are sent to the path planning layer, including: Obtain the logical dependencies and execution priorities between each of the meta-tasks, and then determine the task execution order of several meta-tasks based on the logical dependencies and execution priorities; In response to any abnormal execution result of the meta-task, the emergency response plan task corresponding to the abnormal execution result is obtained, and then the task execution order is updated based on the emergency response plan task; A robotic arm is assigned to each of the meta-tasks according to the task execution order, and several of the meta-tasks are sent to the path planning layer.

Images