Turbine rotor blade intelligent assembly method based on digital twinning and reinforcement learning

Abstract

The application relates to a kind of intelligent assembly methods of steam turbine rotor blade based on digital twinning and reinforcement learning, comprising: creating digital twinning virtual space corresponding to physical environment;Static performance data is written in virtual space, and dynamic data is collected and stored in database;Establish path planning and optimization algorithm library, different planning paths are obtained through various path planning and optimization algorithms;Establish reinforcement learning prediction algorithm library, predict the success rate of the current path in the current blade assembly process through different prediction algorithms, determine the running path;Mechanical arm assembles according to running path, and after assembly is completed, whether blade assembly position meets standard is detected according to blade position and blade slot position, and reinforcement learning target detection algorithm library is established;If it meets the standard, the next blade is assembled, otherwise, the reason that does not meet the standard is determined and optimized.Compared with prior art, the application has the advantages of reducing assembly failure rate, high assembly efficiency and the like.

Description

Technical Field

[0001] This invention relates to the field of intelligent assembly technology, and in particular to an intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning. Background Technology

[0002] A robotic arm is a complex system with advantages such as high precision, high nonlinearity, strong coupling, and high flexibility, but it also suffers from uncertainties such as parameter perturbations, external interference, and unmodeled dynamics. Robotic arms are widely used in manufacturing to reduce labor costs and improve production efficiency.

[0003] Currently, the assembly of rotor blades using robotic arms relies on traditional techniques that involve manual operation of the robotic arm's posture and movement. This type of control method has limited intelligence. When the robotic arm system automates the assembly of rotor blades, it can only follow a pre-set trajectory. However, the simplistic algorithm used to generate this trajectory significantly impacts the robotic arm's flexibility. The effectiveness of the robotic arm's control directly affects the accuracy and efficiency of production and research. Therefore, achieving intelligent control of the robotic arm is crucial.

[0004] Digital twins combine technologies such as 3D modeling, communication connectivity, and big data analytics to map entities in the physical environment onto a virtual environment. Through simulation technology, the entities are controlled, possible states are predicted, and it is determined whether they can operate accurately in the physical environment. Coordination and unity are achieved through the interaction and fusion of the physical and virtual environments.

[0005] Reinforcement learning is a method in machine learning, a learning mechanism that learns how to map states to behaviors to maximize rewards. Reinforcement learning requires continuous experimentation in an environment, using feedback to continuously optimize the correspondence between states and behaviors. It is commonly used for data analysis, prediction, and object detection. Prediction and object detection algorithms have various branches, which can be trained to obtain efficient and accurate results.

[0006] CN115674191A discloses a robotic arm grasping control method and system based on digital twins, comprising: periodically collecting state information over time during robotic arm operation; converting the state information to construct a digital twin of the robotic arm; training the constructed digital twin of the robotic arm using a deep reinforcement learning algorithm; analyzing and storing the state information of the robotic arm during training; obtaining the optimal path planning after training; analyzing potential fault points in the robotic arm control process and comparing them with historical data results to obtain optimal path planning data; sending the optimal path planning data to the robotic arm and controlling its operation; repeating the above steps until the robotic arm accurately reaches the target point and completes the grasping action. However, the above method only uses one path planning algorithm, and the obtained path may not be optimal. Furthermore, if applied to the field of blade assembly, this method lacks the detection of the robotic arm control results to determine whether the assembly is accurately completed. Therefore, current turbine rotor blade assembly methods suffer from defects such as high assembly error rate and low assembly efficiency. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent assembly method for turbine rotor blades based on digital twins and reinforcement learning. This method optimizes the assembly path while detecting assembly qualification, reducing the assembly error rate. At the same time, it uses reinforcement learning algorithms to predict the assembly success rate and perform qualification detection, thereby improving assembly efficiency.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] A smart assembly method for steam turbine rotor blades based on digital twins and reinforcement learning includes the following steps:

[0010] S1: Create a digital twin virtual space corresponding to the turbine rotor blade assembly robot arm in the physical environment;

[0011] S2: Write the static performance data of the robotic arm into the virtual space and collect the dynamic data of each run;

[0012] S3: Create a database to store the static performance data written and the dynamic data from each run.

[0013] S4: Establish a path planning and optimization algorithm library and use the data stored in the database to train the algorithm. Write in static performance data, the current position of the robotic arm and the blade assembly position. Calculate different planning paths through various path planning and optimization algorithms in the algorithm library, and simulate the planning path of the robotic arm when assembling the blade in virtual space.

[0014] S5: Establish a reinforcement learning prediction algorithm library and use the data stored in the database to train the algorithm. Write the planned path and use different reinforcement learning prediction algorithms, combined with the static performance data of the robotic arm, to predict the success rate of the current planned path in this blade assembly process. If the average success rate of the planned path is greater than the threshold, take the planned path with the largest average success rate as the running path in the physical environment. Otherwise, return to the path planning and optimization algorithm library to optimize the path and re-plan the path.

[0015] S6: The robotic arm runs along a determined path in the physical environment. After the operation is completed, it collects the blade position information based on the depth camera at the end of the robotic arm. Combined with the blade slot position information, a reinforcement learning target detection algorithm library is established and the algorithm is trained using the data stored in the database. The trained reinforcement learning target detection algorithm is then used to detect whether the blade assembly position meets the standard.

[0016] S7: If the blade assembly position meets the standard, proceed with the assembly of the next blade; otherwise, transmit the non-standard blade position information back to the digital twin space and determine the reason for non-compliance. Provide optimization suggestions through the algorithm library and assist in adjustment or reassembly through a robotic arm.

[0017] The static performance data of the robotic arm includes its position, posture, grip strength, and service life.

[0018] The path planning and optimization algorithm library includes A*, D*, BFS, and DFS path planning algorithms.

[0019] The reinforcement learning prediction algorithm library includes ANN, Q-learning, KNN, and Boosting prediction algorithms.

[0020] The reinforcement learning object detection algorithm library includes CNN, RCNN, and YOLOv5 object detection algorithms.

[0021] The input to the reinforcement learning target detection algorithm library is the three-dimensional coordinates of the blade slot pixels and the three-dimensional coordinates of the blade pixels.

[0022] Based on the images captured by the depth camera at the end of the robotic arm, the depth camera coordinate system is converted into a pixel coordinate system. The three-dimensional coordinates of the blade groove and the blade in the pixel coordinate system are obtained based on the correspondence between the two-dimensional and three-dimensional images.

[0023] The reasons for non-compliance include misalignment of blade assembly position and blade falling off.

[0024] When the reason for non-compliance with the standard is that the blade assembly position is off, it is determined whether the offset angle affects the use. If it does, the path is adjusted by planning and optimizing the algorithm library, and a robotic arm is used to assist in adjusting the blade back to the designated position.

[0025] When the reason for non-compliance is that the blade has fallen off, the rotor is reset to the installation position of the empty blade slot and the blade is reassembled.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) This invention uses reinforcement learning prediction algorithm and target detection algorithm in digital twin virtual space to optimize the assembly path and detect the assembly qualification, thereby effectively reducing the error rate of blade assembly.

[0028] (2) The present invention uses digital twin technology to provide a virtual environment for displaying the form of the robotic arm operation, thereby reducing the cost of implementation.

[0029] (3) This invention combines path planning algorithm, prediction algorithm and target detection algorithm. The prediction algorithm can effectively predict the assembly success rate of different paths. The target detection algorithm can detect whether the blade assembly position is qualified and assist the robotic arm in adjusting the target. The reinforcement learning algorithm can improve assembly efficiency and obtain the optimal assembly path.

[0030] (4) The present invention can complete the blade assembly work independently for a long time, which can not only improve the assembly efficiency, but also reduce labor costs. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the interaction between the virtual and physical environments corresponding to this invention.

[0033] Figure 3 This is a schematic diagram of the blade assembly logic framework of the method of the present invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0035] This embodiment provides an intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning, such as... Figure 1 As shown, it includes the following steps:

[0036] S1: Connect the rotor blade assembly unit and robotic arm in the physical environment to the PLC communication network and sensors, and connect the other end to the digital twin data interface to parse the data, then render the scene, and create a digital twin virtual space corresponding to the turbine rotor blade assembly robotic arm in the physical environment through 3D modeling software. The two environments correspond to each other.

[0037] A schematic diagram illustrating the interaction between the virtual and physical environments is shown below. Figure 2 As shown, the physical environment feeds back relevant information about the robotic arm's operation to the virtual environment, which then guides the actual operation of the robotic arm in the physical environment based on the simulation results. The virtual environment includes databases (static and dynamic), a path planning and optimization algorithm library, a reinforcement learning prediction algorithm library, and a reinforcement learning target detection algorithm library, which will be described in detail below.

[0038] S2: Write the static performance data of the robotic arm into the virtual space, including the position, posture, grip strength, and service life of the robotic arm; and collect dynamic data for each operation.

[0039] Specifically, static performance data such as the robotic arm's position and posture, and grip strength are extracted. This data is preprocessed and cleaned before being written into a digital twin system and applied to the blade assembly robotic arm in the virtual environment. The robotic arm's static performance data needs to be collected at regular intervals to record wear and tear from repeated use, reducing errors caused by details such as insufficient grip strength.

[0040] It is necessary not only to collect static data of the robotic arm, but also to collect dynamic data of the robotic arm during each operation via PLC communication, so as to help the reinforcement learning algorithm to make continuous corrections and thus improve accuracy.

[0041] S3: Create a database to store the static performance data written and the dynamic data from each run.

[0042] Databases can be divided into two categories: static databases and dynamic databases. Static databases store robotic arm performance information, including data collection and coverage for each time period to understand the degree of wear and tear. Dynamic databases store dynamic data from each operation, such as path data and pass / fail detection data, facilitating data extraction by algorithm libraries.

[0043] Because the robotic arm is used for a long time, the dynamic database collects a large amount of data in real time. While the static database collects current performance information at each time period, the dynamic database can delete dynamic information from a period of time in order to ensure database storage efficiency.

[0044] S4: Establish a path planning and optimization algorithm library and use the data stored in the database to train the algorithm. Write in static performance data, the current position of the robotic arm and the blade assembly position. Calculate different planning paths through various path planning and optimization algorithms in the algorithm library, and simulate the planning path of the robotic arm when assembling the blade in virtual space.

[0045] In this embodiment, the path planning and optimization algorithm library includes path planning algorithms such as A*, D*, BFS, and DFS. First, historical runtime data is collected. Then, the algorithm logic is developed on the development platform, and the historical data is transferred to the development platform as a training set for iterative training of the model algorithm. The algorithm logic source code is packaged and input into the digital twin system, where the algorithm is dynamically compiled in a virtual environment, completing the initial editing of the algorithm library.

[0046] The path planning and optimization algorithm library first requires a password to access the static database, then the blade assembly position needs to be determined using a depth camera, and the current position of the robotic arm needs to be clear. Using this data from three aspects, the algorithm in the library is invoked for path planning. Different algorithms will produce different paths, and all different paths will be transmitted to the reinforcement learning prediction algorithm library in step S5.

[0047] S5: Establish a reinforcement learning prediction algorithm library and use the data stored in the database to train the algorithm. Write the planned path and use different reinforcement learning prediction algorithms, combined with the static performance data of the robotic arm, to predict the success rate of the current planned path in this blade assembly process. If the average success rate of the planned path is greater than the threshold, take the planned path with the largest average success rate as the running path in the physical environment. Otherwise, return to the path planning and optimization algorithm library to optimize the path and re-plan the path.

[0048] In this embodiment, the reinforcement learning prediction algorithm library includes ANN, Q-learning, KNN, and Boosting prediction algorithms. First, historical runtime data is collected. Then, the algorithm logic is developed on the development platform, and the historical data is transferred to the development platform as a training set for iterative training of the model algorithm. The algorithm logic source code is packaged and input into the digital twin system, where the algorithm is dynamically compiled in a virtual environment, completing the initial editing of the algorithm library.

[0049] In the reinforcement learning prediction algorithm library, various algorithm paths are first received from the path planning and optimization algorithm library. The prediction algorithms in the reinforcement learning prediction algorithm library are called. Each path will be predicted by all prediction algorithms to obtain the prediction success rate. The average of all prediction success rates is calculated, and the path with the higher average is the actual path to be run.

[0050] Before building a reinforcement learning prediction algorithm library, it is necessary to train the prediction algorithm with past data. After training, the reinforcement learning algorithm can make predictions more accurately.

[0051] In a digital twin system, simulation is used to determine whether it can be used on a physical entity. A reinforcement learning prediction algorithm library is used to receive different paths for this operation, extract static performance data of the robotic arm, and derive the success rate of each path based on different prediction algorithms.

[0052] Take the average of all prediction algorithms for each path. If the average prediction is greater than 90%, select the path with the highest average as the running path in the physical environment. If the average prediction is less than 90%, write the prediction model into the path planning and optimization algorithm library, optimize the path, and replan.

[0053] For example, in a path planning and optimization algorithm library, four algorithms—A*, D*, BFS, and DFS—are used to derive four different paths: A* path, D* path, BFS path, and DFS path. These four paths are then transmitted to a reinforcement learning prediction algorithm library. In this library, using four prediction algorithms—ANN, Q-learning, KNN, and Boosting—the A* path will obtain four prediction success rates from these four algorithms. These four success rates are then averaged to obtain the average success rate of the A* path. Similarly, the D*, BFS, and DFS paths will also obtain four prediction success rates from these four algorithms, and their average success rates will be calculated. In other words, the reinforcement learning prediction algorithm library outputs the success rate of the squared sum of the number of prediction algorithms for the number of paths, ultimately yielding the average success rate for the total number of paths. Finally, the higher average success rate is selected as the actual running path.

[0054] S6: The robotic arm runs along a determined path in the physical environment. After the operation is completed, it collects the blade position information based on the depth camera at the end of the robotic arm. Combined with the blade slot position information, a reinforcement learning target detection algorithm library is established and the algorithm is trained using the data stored in the database. The trained reinforcement learning target detection algorithm is then used to detect whether the blade assembly position meets the standard.

[0055] In this embodiment, the reinforcement learning object detection algorithm library includes CNN, RCNN, and YOLOv5 object detection algorithms. First, historical runtime data is collected. Then, the algorithm logic is developed on the development platform, and the historical data is transferred to the development platform as a training set for iterative training of the model algorithm. The algorithm logic source code is packaged and input into the digital twin system, where the algorithm is dynamically compiled in a virtual environment, completing the initial editing of the algorithm library.

[0056] Once the blade assembly robot arm is running in the physical environment, a depth camera fixed to the end of the robot arm takes pictures of the assembled blades, reads the blade slot and blade coordinates, and inputs them into a reinforcement learning target detection algorithm.

[0057] In the reinforcement learning-based object detection algorithm library, the blade position information retrieved from the depth camera at the end of the robotic arm is first received. This information is then matched with the blade slot position information, and algorithms from the library are used to determine whether the blade position on the slot conforms to a standard. All object detection algorithms in the library are used for testing to prevent errors from a single algorithm.

[0058] Specifically, based on the images captured by the depth camera at the end of the robotic arm, the depth camera coordinate system is converted into a pixel coordinate system. Based on the correspondence between the two-dimensional and three-dimensional images, the three-dimensional coordinates (x1, y1, z1) of the blade groove pixels and the three-dimensional coordinates (x2, y2, z2) of the blade pixels on the blade groove are obtained.

[0059] The three-dimensional coordinates (x1, y1, z1) of the blade slot pixel and the three-dimensional coordinates (x2, y2, z2) of the blade pixel are written into the target detection algorithm library, and the target detection algorithm is used to detect whether the assembly position meets the standard.

[0060] S7: If the blade assembly position meets the standard, the information is fed back to the digital twin system to assemble the next blade. Otherwise, the non-standard blade position information is transmitted back to the digital twin space and the reason for non-compliance is determined by the target detection algorithm. The algorithm library provides optimization suggestions and the robotic arm assists in adjustment or reassembly.

[0061] The reasons for non-compliance include misalignment of the blade assembly position and blade falling off.

[0062] When the reason for non-compliance is that the blade assembly position is off, the offset information is fed back to the digital twin system to determine whether the offset angle will affect the assembly of the next blade and the use of the rotor. If it will affect the use, the path will be adjusted by planning and optimizing the algorithm library, and the robotic arm will be controlled to make auxiliary adjustments. After the adjustment is completed, the target algorithm will be used again to detect whether the assembly is qualified.

[0063] When the reason for non-compliance is blade loss, the blade loss information is fed back to the digital twin system, the rotor is reset to the installation position of the empty blade slot, and the blade assembly is repeated.

[0064] like Figure 3As shown, this invention utilizes digital twin technology to construct a virtual environment corresponding to the physical environment of the robotic arm blade. The two environments are interconnected, and the parameter data of the robotic arm is written into the virtual space database. Through simulation in the virtual space, the path of the robotic arm in the physical environment is planned. After the path is obtained, simulation is performed to check for problems and whether it can be applied to a physical entity. Using a reinforcement learning prediction algorithm, based on the training set and historical accuracy information, it predicts whether the current path can complete the assembly of the blade. If the prediction success rate is above 90%, it runs in the physical environment. If the prediction result is below 90%, path optimization is performed, the robotic arm path is replanned, and then it runs. After the robotic arm assembles the blade, a reinforcement learning target detection algorithm, in conjunction with a depth camera at the end of the robotic arm, is used to calculate and detect whether the blade assembly position meets the standards. If it is qualified, the assembly of the next blade can proceed; if it is not qualified, the reason for the failure is determined, such as position deviation or falling, and then fed back to the digital twin virtual environment. If it is due to position deviation, it is determined whether the position deviation angle affects the use of the blade. If it does, the robotic arm assists in adjustment; if it is due to falling, the blade assembly task is repeated. This invention optimizes and predicts the assembly path of robotic arm blades, reducing the probability of assembly errors and improving assembly efficiency. On the other hand, it enables position detection of blade assembly, preventing blade position deviation or falling from affecting subsequent assembly and the use of the rotor disk.

[0065] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for intelligent assembly of steam turbine rotor blades based on digital twins and reinforcement learning, characterized in that, Includes the following steps: S1: Create a digital twin virtual space corresponding to the turbine rotor blade assembly robot arm in the physical environment; S2: Write the static performance data of the robotic arm into the virtual space and collect the dynamic data of each run; S3: Create a database to store the static performance data written and the dynamic data from each run. S4: Establish a path planning and optimization algorithm library and use the data stored in the database to train the algorithm. Write in static performance data, the current position of the robotic arm and the blade assembly position. Calculate different planning paths through various path planning and optimization algorithms in the algorithm library, and simulate the planning path of the robotic arm when assembling the blade in virtual space. S5: Establish a reinforcement learning prediction algorithm library and use the data stored in the database to train the algorithm. Write the planned path and use different reinforcement learning prediction algorithms, combined with the static performance data of the robotic arm, to predict the success rate of the current planned path in this blade assembly process. If the average success rate of the planned path is greater than the threshold, take the planned path with the largest average success rate as the running path in the physical environment. Otherwise, return to the path planning and optimization algorithm library to optimize the path and re-plan the path. S6: The robotic arm runs along a determined path in the physical environment. After the operation is completed, it collects the blade position information based on the depth camera at the end of the robotic arm. Combined with the blade slot position information, a reinforcement learning target detection algorithm library is established and the algorithm is trained using the data stored in the database. The trained reinforcement learning target detection algorithm is then used to detect whether the blade assembly position meets the standard. S7: If the blade assembly position meets the standard, proceed with the assembly of the next blade; otherwise, transmit the non-standard blade position information back to the digital twin space and determine the reason for non-compliance. Provide optimization suggestions through the algorithm library and assist in adjustment or reassembly through a robotic arm.

2. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 1, characterized in that, The static performance data of the robotic arm includes its position, posture, grip strength, and service life.

3. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 1, characterized in that, The path planning and optimization algorithm library includes A*, D*, BFS, and DFS path planning algorithms.

4. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 1, characterized in that, The reinforcement learning prediction algorithm library includes ANN, Q-learning, KNN, and Boosting prediction algorithms.

5. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 1, characterized in that, The reinforcement learning target detection algorithm library includes CNN, RCNN, and YOLOv5 target detection algorithms.

6. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 1, characterized in that, The input to the reinforcement learning target detection algorithm library is the three-dimensional coordinates of the blade slot pixels and the three-dimensional coordinates of the blade pixels.

7. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 6, characterized in that, Based on the images captured by the depth camera at the end of the robotic arm, the depth camera coordinate system is converted into a pixel coordinate system. The three-dimensional coordinates of the blade groove and the blade in the pixel coordinate system are obtained based on the correspondence between the two-dimensional and three-dimensional images.

8. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 1, characterized in that, The reasons for non-compliance include misalignment of the blade assembly position and blade falling off.

9. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 8, characterized in that, When the reason for non-compliance with the standard is that the blade assembly position is off, it is determined whether the offset angle affects the use. If it does, the path is adjusted by planning and optimizing the algorithm library, and a robotic arm is used to assist in adjusting the blade back to the designated position.

10. The intelligent assembly method for steam turbine rotor blades based on digital twins and reinforcement learning according to claim 8, characterized in that, When the reason for non-compliance is that the blade has fallen off, the rotor is reset to the installation position of the empty blade slot and the blade is reassembled.

Images