A robot assembly optimization method and system based on federated learning

Abstract

The application discloses a robot assembly optimization method and system based on federated learning. The method comprises the following steps: acquiring image data and point cloud data of a robot and performing data feature fusion and recognition processing to obtain comprehensive feature data of the robot; acquiring force sensation information and visual information of the robot, and combining the comprehensive feature data of the robot to perform motion error optimization processing based on deep reinforcement learning, thereby obtaining comprehensive feature data of the robot after error compensation; performing model training on the robot through a federated learning optimization algorithm, thereby constructing a robot with a global assembly control model; and based on a meta-learning strategy, performing a communication cooperation on the robot with the global assembly control model to execute an assembly task, thereby obtaining a robot assembly optimization result. The embodiment of the application can improve the environmental perception capability of the robot, effectively improve the response speed and adaptability of the robot in a multi-task switching scene. The application can be widely applied to the technical field of robot assembly.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot assembly, in particular to a robot assembly optimization method and system based on federated learning. BACKGROUND

[0002] With the rapid development of robot technology, industrial robots are increasingly widely used in precision assembly. Especially in the process of assembling small parts, robots need to have high precision and flexibility to cope with complex and variable environments and assembly requirements. However, most of the current robot assembly systems rely on a single sensor to perceive the assembly environment. Due to the limitations of different sensors in obtaining information, the use of a single sensor is easily disturbed by environmental light, reflection, shielding and other factors, resulting in insufficient environmental perception accuracy. Even if some solutions introduce sensor fusion technology, due to the complexity and imperfection of the data fusion algorithm, the accuracy and robustness of the fused information are still difficult to fully guarantee. In the process of precision assembly, error compensation in the existing assembly system is usually based on simple feedback control rather than complex real-time learning models. The feedback mechanism is lagging, and the error compensation effect is limited. Especially in dynamic environments or complex tasks, the system is difficult to quickly and adaptively adjust and optimize. The current deep learning models applied to robot precision assembly are mostly trained in offline environments. Each robot may face different assembly tasks and environmental conditions, and a single model is difficult to adapt to the needs of all scenarios. And once the model is deployed, it is difficult to update, which makes the robot unable to learn and adapt to new situations in real time in actual applications. In addition, robots lack effective knowledge sharing and collaboration mechanisms, and each robot can only rely on its own data and experience for operation, resulting in low learning efficiency and poor task execution effect. Related federated learning algorithms mostly rely on synchronous update strategies, and there are still delays in processing large-scale, multi-modal perception data, which lacks real-time performance. In addition, it is easy to be affected by problems such as Byzantine attack, data pollution and lack of sufficient security protection mechanism, which can easily lead to data leakage or system crash.

[0003] In summary, the technical problems in the related art need to be improved. SUMMARY

[0004] The main purpose of the embodiments of the present application is to provide a robot assembly optimization method and system based on federated learning, which can improve the environmental perception ability of robots and effectively improve the response speed and adaptability of robots in multi-task switching scenarios.

[0005] To achieve the above purpose, one aspect of an embodiment of the present application provides a robot assembly optimization method based on federated learning, which comprises:

[0006] Acquire the robot's image data and point cloud data and perform data feature fusion and recognition processing to obtain the robot's comprehensive feature data;

[0007] Acquiring force information and visual information of the robot, and combining the robot's comprehensive feature data with motion error optimization processing based on deep reinforcement learning to obtain error-compensated robot comprehensive feature data;

[0008] Based on the comprehensive feature data of the robot after error compensation, the robot model is trained through a federated learning optimization algorithm to construct a robot with a global assembly control model;

[0009] Based on the meta-learning strategy, the robots with the global assembly control model communicate and cooperate to perform assembly tasks, and obtain robot assembly optimization results.

[0010] In some embodiments, the acquiring of the robot's image data and point cloud data and performing data feature fusion and recognition processing to obtain the robot's comprehensive feature data includes:

[0011] Calibrate the camera and lidar to obtain a calibrated camera and a calibrated lidar;

[0012] Acquiring image data of the robot through the calibrated camera and acquiring point cloud data of the robot through the calibrated lidar;

[0013] Performing data preprocessing and data fusion processing on the image data of the robot and the point cloud data of the robot to obtain fused robot data;

[0014] Introducing image branch module, point cloud branch module and feature fusion layer to build a deep learning fusion model;

[0015] Based on the deep learning fusion model, feature recognition and fusion processing are performed on the fused robot data to obtain the comprehensive feature data of the robot.

[0016] In some embodiments, the acquiring of the robot's force information and the robot's visual information, and combining the robot's comprehensive feature data with motion error optimization processing based on deep reinforcement learning to obtain error-compensated robot comprehensive feature data, includes:

[0017] Acquiring force information of the robot through a force sensor and acquiring visual information of the robot through a visual sensor;

[0018] Defining a state space for deep reinforcement learning based on the comprehensive feature data of the robot, defining an action space for deep reinforcement learning based on the force information and visual information of the robot, setting an experience replay pool, and conducting interactive learning to obtain preliminarily optimized comprehensive feature data of the robot;

[0019] Controlling the robot to perform simulation according to the preliminary optimized comprehensive characteristic data of the robot to obtain a robot motion deviation vector;

[0020] The robot motion deviation vector is error compensated by an error compensation algorithm, and is optimized in combination with a deep reinforcement learning algorithm to obtain the comprehensive feature data of the robot after error compensation.

[0021] In some embodiments, the method of performing model training on the robot using a federated learning optimization algorithm based on the error-compensated comprehensive feature data of the robot to construct a robot with a global assembly control model includes:

[0022] Controlling the robot to perform an assembly task according to the error-compensated comprehensive feature data of the robot, and constructing a local data set according to the assembly execution result;

[0023] Performing data preprocessing on the local data set to obtain a preprocessed local data set;

[0024] A robot assembly control model is constructed based on the multi-sensor fusion model and the motion control model, and trained using a small-batch gradient descent algorithm to obtain the trained robot assembly control model.

[0025] The pre-processed local data set is compressed and uploaded to the trained robot assembly control model through the AES symmetric encryption algorithm;

[0026] Based on the trained robot assembly control model, assigning weight parameters to the preprocessed local data set and performing aggregation processing through an asynchronous federated averaging algorithm to obtain a global assembly control model;

[0027] The global assembly control model is sent to a robot control terminal to obtain the robot having the global assembly control model.

[0028] In some embodiments, the robot having the global assembly control model communicates and collaborates to perform assembly tasks based on a meta-learning strategy to obtain a robot assembly optimization result, including:

[0029] Based on an industrial communication protocol, the robots having the global assembly control model are networked for communication;

[0030] Introducing a personalized network model, performing global network fine-tuning processing on the robot having the global assembly control model, and obtaining a fine-tuned robot;

[0031] A meta-learning strategy is introduced to perform adaptive training on the fine-tuned robot to obtain an adaptively trained robot;

[0032] Based on a continuous learning mechanism, dynamically adjusting the adaptively trained robot to obtain a robot with multi-task compatibility;

[0033] The robot assembly optimization result is obtained by executing the assembly task based on the robot with multi-task compatibility.

[0034] In some embodiments, the personalized network model includes several shared layers and several dedicated layers, and output ends of several shared layers are connected to input ends of several dedicated layers, wherein:

[0035] The shared layers are used to extract common features of the robot;

[0036] The dedicated layers are used to extract the personalized demand features of the robot.

[0037] In some embodiments, the introducing of a meta-learning strategy to adaptively train the fine-tuned robot to obtain an adaptively trained robot includes:

[0038] Initializing the fine-tuned robot using a meta-learning algorithm to obtain an initialized robot;

[0039] The end control and force feedback parameters of the initialized robot are adjusted to obtain the adaptively trained robot.

[0040] In some embodiments, the method of dynamically adjusting the adaptively trained robot based on a continuous learning mechanism to obtain a robot with multi-task compatibility includes:

[0041] Based on a continuous learning mechanism, dynamically adjusting the learning rate of the adaptively trained robot to obtain an adjusted robot;

[0042] The adjusted robot is subjected to inter-model knowledge transfer learning through a knowledge distillation algorithm to obtain the robot with multi-task compatibility.

[0043] In some embodiments, the method further includes adding noise during the process of communicating and cooperating with the robots having the global assembly control model to perform assembly tasks.

[0044] To achieve the above object, another aspect of the embodiment of the present application proposes a robot assembly optimization system based on federated learning, comprising:

[0045] A first module is configured to acquire image data and point cloud data of the robot and perform data feature fusion and recognition processing to obtain comprehensive feature data of the robot;

[0046] A second module is configured to acquire force sensation information of the robot and visual information of the robot, and perform motion error optimization processing based on deep reinforcement learning in combination with the comprehensive feature data of the robot to obtain comprehensive feature data of the robot after error compensation;

[0047] A third module is configured to perform model training on the robot based on the comprehensive feature data of the robot after error compensation through a federated learning optimization algorithm to construct a robot with a global assembly control model;

[0048] A fourth module is configured to perform an assembly task in communication and cooperation based on a meta-learning strategy for the robot with the global assembly control model to obtain a robot assembly optimization result.

[0049] The embodiment of the present application at least has the following beneficial effects: the present application provides a robot assembly optimization method and system based on federated learning, which acquires image data and point cloud data of the robot and performs data feature fusion and recognition processing, uses multi-sensor fusion and deep learning control algorithm to enable the robot to have high-precision environmental perception ability in complex assembly tasks, further acquires force sensation information of the robot and visual information of the robot, and performs motion error optimization processing based on deep reinforcement learning in combination with the comprehensive feature data of the robot, detects assembly errors and performs automatic compensation through real-time feedback of force sensation sensors and visual sensors, and the error detection module combines force sensation and visual information to calculate a deviation vector in real time. The compensation algorithm quickly corrects the motion path through the controller to ensure high-precision operation of the robot in the assembly process. The error compensation module predicts historical error patterns through deep learning to make adjustments in advance, improves the speed and accuracy of compensation, further performs model training on the robot based on the comprehensive feature data of the robot after error compensation through a federated learning optimization algorithm, realizes distributed model training and sharing of each robot through federated learning under a multi-robot cooperation framework, and finally performs an assembly task in communication and cooperation based on a meta-learning strategy for the robot with the global assembly control model to train the robot model to have rapid adaptation ability, so that it can quickly adjust to new task requirements through a small number of training steps when the task is switched or the environment is changed, and effectively improves the response speed and adaptability of the system in a multi-task switching scenario. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1This is a flowchart of a robot assembly optimization method based on federated learning provided in an embodiment of the present application;

[0051] Figure 2 This is a structural diagram of a robot assembly optimization system based on federated learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are merely examples of systems and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.

[0053] It will be understood that the terms "first", "second", etc. used in this application may be used herein to describe various concepts, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0054] The terms "at least one", "plurality", "each", "any", etc. used in this application include "at least one", "two" or more, "plurality" or "each", "any" or "any one", "each" or "any one" as used herein.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0056] There are some deficiencies in the related technology, such as:

[0057] 1) Reliance on a single sensor and insufficient perception capabilities.

[0058] 2) Lack of efficient error compensation mechanism;

[0059] 3) Limitations of deep learning models and insufficient localized adaptability.

[0060] 4) Multi-robot collaboration is inefficient and lacks an effective knowledge sharing mechanism.

[0061] 5) Real-time and asynchronous issues in federated learning applications.

[0062] 6) Security and reliability issues.

[0063] In view of this, the present application provides a federated learning-based robot assembly optimization method. By distributing training and aggregating models among nodes (robots) without sharing data, this method can protect data privacy while enhancing the model's generalization ability and learning effect. In the robotic precision assembly scenario, federated learning can utilize the perception and assembly data of different robots in their respective working environments to optimize the model, thereby improving the accuracy, real-time performance, and adaptability of the overall assembly.

[0064] Reference Figure 1 , Figure 1 A flowchart of a robot assembly optimization method based on federated learning provided by an embodiment of the present invention, referring to Figure 1 , the method comprises the following steps:

[0065] S100, acquiring the robot's image data and point cloud data and performing data feature fusion and recognition processing to obtain the robot's comprehensive feature data;

[0066] It should be noted that, in some embodiments, step S100 may include steps S110 to S150.

[0067] S110, performing calibration processing on the camera and the laser radar to obtain a calibrated camera and a calibrated laser radar;

[0068] In step S110 of some embodiments, for example, the sensor is calibrated, and the internal and external parameters of the camera are calibrated using a calibration plate or a specific marker to obtain the intrinsic parameter matrix and distortion parameters of the camera. The lidar is calibrated to obtain the conversion relationship between its coordinate system and the robot coordinate system. The precise coordinate transformation matrix between the camera, lidar and the robot end effector is determined through the hand-eye calibration method.

[0069] S120, acquiring image data of the robot through the calibrated camera and acquiring point cloud data of the robot through the calibrated lidar;

[0070] In step S120 of some embodiments, illustratively, the camera and the lidar start data acquisition simultaneously, the camera captures image frames, the lidar scans to obtain point cloud data, and the camera image and lidar data are synchronized through timestamps to ensure the temporal consistency of multi-sensor data.

[0071] S130, performing data preprocessing and data fusion processing on the robot's image data and the robot's point cloud data to obtain fused robot data;

[0072] In step S130 of some embodiments, the image data is preprocessed, illustratively, by performing operations such as image denoising, grayscale conversion, and enhancement to improve image quality. ROI (region of interest) extraction is used to focus on the assembly area to reduce the amount of computation, and the point cloud data is then preprocessed by downsampling and filtering to remove noise and outliers. Based on the assembly task, key point cloud areas are extracted, and finally spatial fusion is performed. Using known sensor calibration parameters, the point cloud data of the lidar is projected into the image coordinate system of the camera, or the image data is mapped into the point cloud coordinate system to achieve spatial alignment.

[0073] S140, introduce the image branch module, point cloud branch module and feature fusion layer to build a deep learning fusion model;

[0074] S150. Perform feature recognition and fusion processing on the fused robot data based on the deep learning fusion model to obtain comprehensive feature data of the robot.

[0075] In steps S140 to S150 of some embodiments, exemplarily, a fusion convolutional neural network (Fusion-CNN) is used, which consists of an image processing branch and a point cloud processing branch. The image branch uses classic network structures such as ResNet and VGG to extract image features. The point cloud branch uses networks such as PointNet and PointCNN to extract point cloud features. The feature fusion layer fuses the image features and the point cloud features to generate a comprehensive feature representation, and then constructs a training set containing multimodal data, with labels such as the position, posture, category, and other information of the assembly object. Supervised learning methods are used to minimize the loss function (such as cross entropy and mean square error) between the predicted results and the true labels. Data enhancement technology is used to improve the robustness of the model, generate high-precision position information and posture information of the assembly object, and detection results of environmental obstacles, and provide data support for subsequent motion planning.

[0076] S200, acquiring force information and visual information of the robot, and performing motion error optimization processing based on deep reinforcement learning in combination with comprehensive feature data of the robot to obtain comprehensive feature data of the robot after error compensation;

[0077] It should be noted that, in some embodiments, step S200 may include steps S210 to S240.

[0078] S210, acquiring force information of the robot through a force sensor and acquiring visual information of the robot through a visual sensor;

[0079] In step S210 of some embodiments, for example, a force sensor is installed on the robot's end effector, capable of monitoring the forces and torques applied during the assembly process in real time. When an abnormal force change is detected, the system determines that an assembly error or collision may have occurred. Through force feedback, the system can capture subtle mechanical changes that are imperceptible to the naked eye. This is crucial for error detection in precision assembly. For example, if the position of a part held by an actuator shifts slightly during assembly, the force sensor will sense that the forces acting on the actuator are no longer balanced. The system uses this information to determine the direction and magnitude of the error and compensate for it. A high-resolution camera captures subtle displacement changes during the assembly process in real time. Visual sensors can detect whether the assembly object is in the correct predetermined position and whether its posture is accurate. If the assembly object shifts or has an incorrect posture, the system immediately issues an alarm and records the offset for error analysis and correction. For example, during precision assembly, the image captured by the camera can provide real-time feedback on the alignment of the currently assembled parts. If a slight rotational deviation is detected, the system can issue a corrective instruction through visual feedback.

[0080] S220: defining a state space for deep reinforcement learning based on the robot's comprehensive feature data, defining an action space for deep reinforcement learning based on the robot's force information and visual information, setting an experience replay pool, and conducting interactive learning to obtain preliminarily optimized comprehensive feature data for the robot;

[0081] In step S220 of some embodiments, illustratively, a state space is set, which defines the key information that the robot needs to perceive at each moment. In precision assembly tasks, the state space includes the following elements: Robot joint angles and velocities: This information describes the current posture and motion state of the robotic arm. Position and posture of the assembly object: The precise position and posture of the assembly object are obtained through sensors (such as cameras and lidars). Environmental obstacle information: To prevent collisions during the assembly process, the system will identify and record the position information of surrounding obstacles.

[0082] An action space is set up, which defines the specific actions the robot can perform in each state. In this invention, the action space includes: Joint angular velocity instructions: These adjust the angular velocity of each joint of the robot arm to control the robot's motion trajectory. End effector displacement instructions: These directly control the displacement of the robot's end effector in space, ensuring that the actuator accurately reaches the target position. The design of the action space requires a balance between flexibility and control accuracy to ensure that the robot can accurately perform assembly operations in complex environments.

[0083] Setting a reward function is a core component of reinforcement learning, determining how the robot adjusts its actions to achieve optimal task performance. In precision assembly, the reward function is designed based on the following factors: Task completion: Each time the robot successfully completes an assembly task, the system rewards it with a higher reward. Error magnitude: As the robot approaches the target position, the smaller the error, the higher the reward, encouraging the robot to improve assembly accuracy. Negative rewards are set to penalize the robot when the task takes too long or consumes too much energy, thereby guiding it to choose more efficient actions in future operations. The reward function should be designed to consider both assembly accuracy and efficiency, ensuring that the robot completes the task as quickly and efficiently as possible while meeting accuracy requirements. To reduce the risk of exploration in a real environment, the initial model is pre-trained in a simulation environment. Through simulation, the robot can accumulate experience in virtual scenes and learn how to optimize its motion strategy. Specific training algorithms include: Deep Q-Network (DQN): Selects the optimal action in a discrete action space. Proximal Policy Optimization (PPO): Applicable to continuous action spaces, enabling more precise adjustment of the robot's actions. After the training in the simulation phase is completed, fine-tuning is performed based on data from actual assembly tasks to improve the model's adaptability and generalization capabilities, making it more robust in real-world environments.

[0084] S230, controlling the robot to perform simulation based on the preliminarily optimized comprehensive feature data of the robot, and obtaining a robot motion deviation vector;

[0085] In step S230 of some embodiments, illustratively, after error detection, the system first needs to analyze the error. By combining the data of force feedback and visual feedback, the system can calculate the deviation vector between the current position and the target position. The deviation vector not only describes the offset in position, but also includes the deviation in posture (rotation). Through precise error analysis, the system can understand the gap between the current state of the robot and the ideal state, thereby providing a basis for compensation. If the visual feedback detects that the assembly object has a 1mm offset in the XY plane, the force sensor detects that an additional 0.1N torque force occurs during the assembly process. The system will generate a comprehensive deviation vector by combining these two error information to guide subsequent compensation actions.

[0086] S240. Perform error compensation on the robot motion deviation vector through an error compensation algorithm, and optimize it in combination with a deep reinforcement learning algorithm to obtain comprehensive feature data of the robot after error compensation.

[0087] In step S240 of some embodiments, the compensation algorithm illustratively calculates the motion instructions that require adjustment based on the error analysis results. In this system, the compensation algorithm uses a classic PID (proportional-integral-derivative) controller or model predictive control (MPC) to correct the robot's motion in real time. PID controller: A commonly used error correction algorithm, the PID controller calculates output corrections based on the current error (proportional term), the accumulated error (integral term), and the rate of change of the error (derivative term). The advantages of the PID controller lie in its simplicity and ability to correct minor system deviations in real time, making it suitable for most static or slowly changing assembly scenarios. Model predictive control (MPC): Compared to the PID controller, MPC is more intelligent. MPC not only considers the current error but also predicts future error trends by establishing mathematical models of the robot's kinematics and dynamics, allowing it to pre-calculate the optimal control strategy. MPC is suitable for complex assembly tasks, especially in rapidly changing dynamic environments, and can effectively reduce errors and optimize the assembly process. Furthermore, the system incorporates an error compensation module into its deep learning model, leveraging historical error data for learning. By analyzing past assembly data, the system can predict potential error patterns and make appropriate adjustments in advance. For example, if certain assembly tasks are prone to errors in specific steps, the system can make appropriate adjustments in advance to prevent deviations from occurring, and then execute the adjustment strategy. After the error compensation algorithm calculates the corrected motion instructions, the system sends these instructions to the robot controller to adjust the motion trajectory and posture of the robotic arm in real time. In this process, the system ensures that the corrected action is accurate and smooth and does not cause additional interference to the current assembly task. During the adjustment process, the system will also continuously monitor force and visual feedback to ensure the execution effect of the compensation action. For example, when the robot actuator is compensating for the offset action, the force sensor will monitor the force applied by the actuator to ensure that the adjusted force distribution is still balanced. At the same time, the camera will capture the position change of the assembly object to verify whether the adjusted accuracy meets the expected target.

[0088] In summary, the robot's motion control and real-time error correction during the precision assembly process ensure that the robot can complete the assembly task with high precision. During the assembly process, the robot needs to adjust its motion trajectory and posture in real time based on the data fed back by the sensor to adapt to various environmental changes and small displacements of the assembly objects. To this end, an embodiment of the present invention proposes a motion control model based on deep reinforcement learning, and combines force perception with visual feedback to achieve error detection and compensation. In order to enable the robot to perform high-precision operations in a complex assembly environment, a deep reinforcement learning (DRL) strategy is adopted to achieve motion control. The deep reinforcement learning model can continuously optimize its action strategy through interaction with the environment, so that the robot can dynamically adapt to environmental changes when performing tasks.

[0089] S300: Based on the comprehensive feature data of the robot after error compensation, the robot model is trained through a federated learning optimization algorithm to build a robot with a global assembly control model;

[0090] It should be noted that, in some embodiments, step S300 may include steps S310 to S360.

[0091] S310, controlling the robot to perform the assembly task according to the error-compensated comprehensive feature data of the robot, and constructing a local data set according to the assembly execution result;

[0092] In step S310 of some embodiments, illustratively, each robot collects sensor data (such as visual data, lidar data, force data, etc.) and assembly results (including the position, posture, force feedback, etc. of the assembly object) in real time when performing the assembly task. These data constitute a local data set, which reflects the perception and control information of each robot in a specific task. For example, when a robot completes an electronic component assembly task, the collected visual data can record the alignment of the components, and the force data can record the changes in the clamping force. Combined with the success and failure results of the assembly task, these data provide valuable feedback for model training.

[0093] S320, performing data preprocessing on the local data set to obtain a preprocessed local data set;

[0094] In step S320 of some embodiments, the collected data is preprocessed (e.g., through denoising, normalization, feature extraction, etc.) to construct a local dataset suitable for deep learning model training. The labels of the dataset are typically evaluations of the assembly results, such as success or failure, and specific assembly errors.

[0095] S330, constructing a robot assembly control model based on the multi-sensor fusion model and the motion control model, and training the model through a small batch gradient descent algorithm to obtain a trained robot assembly control model;

[0096] In step S330 of some embodiments, illustratively, based on the local data set, the robot trains or fine-tunes the multi-sensor fusion model and the motion control model. The initial model is usually issued by the global server. After fine-tuning with local data, it can better adapt to the unique environment of each robot in a specific task. In order to efficiently update the model parameters, a mini-batch gradient descent (Mini-Batch Gradient Descent) or Adam optimizer is used. These optimization algorithms can converge quickly and adapt to the training of small-scale data sets. For example, using the Adam optimizer, the learning rate can be dynamically adjusted to adapt to data fluctuations in complex assembly scenarios, so that the model can learn key perception and control features faster. Local training is usually performed after each assembly task is completed, or a certain time interval is set according to the task complexity and data volume to ensure that the model can adapt to the new working environment and assembly tasks in a timely manner.

[0097] S340, compressing the pre-processed local data set and uploading it to the trained robot assembly control model through the AES symmetric encryption algorithm;

[0098] In step S340 of some embodiments, illustratively, in order to prevent the model parameters from being intercepted or tampered with during transmission, the system uses the AES symmetric encryption algorithm to encrypt the model parameters. AES is an efficient encryption algorithm that can maintain high transmission efficiency while ensuring data security. In this way, even if the communication network is attacked, the privacy and integrity of the model parameters can be ensured. The data is transmitted through a secure transmission protocol (such as TLS / SSL), further ensuring the security and reliability of the data transmission process. In order to reduce the occupancy of communication bandwidth and speed up the transmission speed, the model parameters are compressed using model compression technology. Common compression methods include weight quantization and model pruning. Weight quantization: converting floating-point weights into low-precision integer representations to reduce the storage and transmission costs of the model. Model pruning: by removing redundant parameters in the model, the size of the model is reduced while maintaining the performance of the model. During the model upload process, the system will dynamically adjust according to the network conditions and bandwidth to ensure that the compressed model parameters can be transmitted to the central server in the shortest possible time.

[0099] S350, based on the trained robot assembly control model, assigning weight parameters to the preprocessed local data set and performing aggregation processing through an asynchronous federated averaging algorithm to obtain a global assembly control model;

[0100] In step S350 of some embodiments, after the central server receives the model updates uploaded by multiple robots, the parameters need to be aggregated to generate a global model. Since robots perform different tasks in different environments, the uploaded model updates may differ, so a reasonable aggregation strategy is needed to optimize the global model. To avoid delays caused by synchronization waiting during transmission, the central server uses the asynchronous federated averaging algorithm (FedAsync). This method allows the server to aggregate immediately after receiving any robot's model update, without waiting for all robot parameters to be complete. This can effectively improve the real-time performance of the system and ensure that the global model can be updated in a timely manner. During the aggregation process, the central server assigns different weights to each robot's model parameters. These weights can be dynamically adjusted based on factors such as the amount of local data, model performance, task importance, etc. For example, data weighting, robots with more local data may contribute more to the model, so their update weight is higher. Performance weighting, according to the robot's assembly success rate, error size, and other indicators to evaluate model performance, the better the model performance, the greater the impact of its update on the global model.

[0101] Further to prevent some robots' model updates from being abnormal (such as due to network errors, device failures or malicious attacks), the server will perform anomaly detection on the received model parameters. By analyzing the model's gradient or parameter update amount, it is detected whether the model deviates from the normal range of global update. For the detected abnormal model, the server will use a method to defend against Byzantine attacks (such as Krum algorithm) for filtering. Krum algorithm selects the part of the model with the smallest update difference for aggregation, preventing malicious or erroneous data from affecting the global model, thereby improving the robustness of the system.

[0102] S360, the global assembly control model is distributed to the robot control terminal to obtain a robot with a global assembly control model.

[0103] In step S360 of some embodiments, after completing model aggregation, the central server needs to distribute the global model to each robot to update their local models, so that each robot shares the latest learning achievements. Similar to model uploading, the server will also encrypt the model parameters when distributing the global model to ensure the security of the data during transmission. AES encryption and TLS / SSL protocol ensure that model parameters cannot be tampered with or intercepted during transmission.

[0104] The local model is then updated. After receiving the global model, the robot needs to fuse it with the local model to maintain a balance between system personalization and global sharing. After receiving the global model, the robot fuses the global model parameters with the local personalized parameters. The first few layers of the model are typically used to extract common features (such as visual features and object recognition). These layers are updated using the global model. The subsequent layers are dedicated to local personalization, adapting to specific task scenarios or assembly conditions. These layers retain the robot's local parameters. Personalized adjustments are made. During the fusion process, the robot can fine-tune the global model appropriately based on its actual task requirements, ensuring that the model both shares global knowledge and adapts to specific local work scenarios. Continuous learning and online updates: After fusing with the global model, the local model maintains its ability to continuously learn. The robot can continue to collect data during assembly tasks and incrementally train the model to ensure that it can continuously adapt to new environments and task changes.

[0105] In summary, the federated learning framework enables continuous optimization of the robot's multi-sensor fusion model and motion control model. The core advantage of federated learning lies in its ability to share model updates across robots through distributed training, eliminating the need to transmit local raw data. This not only protects data privacy but also improves the model's generalization capabilities, enabling it to better adapt to diverse assembly tasks and environmental changes.

[0106] S400, based on the meta-learning strategy, the robots with the global assembly control model communicate and collaborate to perform the assembly task, and obtain the robot assembly optimization result;

[0107] It should be noted that, in some embodiments, step S400 may include steps S410 to S450.

[0108] S410, based on the industrial communication protocol, network communication is performed between robots having the global assembly control model;

[0109] In step S410 of some embodiments, for example, first, the network architecture adopts a distributed network architecture to support direct communication between the robot and the central server, and between the robots. This architecture is based on the TCP / IP protocol stack and combines industrial communication protocols (such as OPC UA and EtherCAT) to achieve highly reliable communication. The security mechanism uses the SSL / TLS protocol to establish a secure communication channel, perform identity authentication and permission management, and prevent unauthorized access.

[0110] Through the central dispatch system, assembly tasks are intelligently allocated according to the status, load and capabilities of each robot. Robots share environmental information, assembly progress, fault conditions, etc. to coordinate work. In the scenario of multi-robot collaboration, collision avoidance algorithms are used to prevent physical conflicts between robots.

[0111] S420, introducing a personalized network model, performing global network fine-tuning processing on the robot having the global assembly control model, and obtaining a fine-tuned robot;

[0112] In step S420 of some embodiments, personalized models and meta-learning are, for example, key modules for improving a robot's adaptability and rapid response capabilities in multiple tasks and environments. In a federated learning framework, while a global model can provide shared foundational knowledge for all robots, personalized model fine-tuning is crucial because each robot faces unique task scenarios and environmental conditions. By combining this with a meta-learning strategy, robots can quickly adapt to new tasks after receiving the global model, while maintaining continuous learning capabilities to cope with dynamically changing working environments.

[0113] To balance global knowledge sharing and local task adaptability, personalized models typically adopt a hierarchical structure. In this design, the first few layers of the model are shared layers, which focus on extracting common features, such as visual feature extraction and basic motion patterns, that are common across different tasks. The shared layers are updated by the global model of federated learning, ensuring that the robot learns common perception and control knowledge from other robots. The latter layers of the model are specialized layers, which focus on addressing the personalized needs of the robot in a specific task or environment. For example, for different assembly tasks, certain robots may need to handle different object types, poses, or constraints. These specialized layers are trained and fine-tuned based on each robot's local data to ensure that the model can well adapt to the unique characteristics of the local task.

[0114] When updating a model, the system prioritizes updating the shared layer to ensure that all robots can utilize globally optimized common features. For the specialized layer, the robot retains or fine-tunes parameters based on the specific circumstances of the local task. This approach enhances personalized processing capabilities while maintaining the overall consistency of the model. For example, for a robot performing an electronic component assembly task, its shared layer is primarily responsible for basic visual feature extraction and posture recognition, while the specialized layer is optimized based on the specific electronic component type and assembly location requirements of the robot to ensure precise execution of the task.

[0115] S430, introducing a meta-learning strategy to perform adaptive training on the fine-tuned robot to obtain an adaptively trained robot;

[0116] Specifically: S431, the fine-tuned robot is initialized through a meta-learning algorithm to obtain an initialized robot; S432, the end control and force feedback parameters of the initialized robot are adjusted to obtain an adaptively trained robot.

[0117] In step S430 of some embodiments, illustratively, meta-learning is a learning method for training models to quickly adapt to new tasks. Unlike traditional training methods, meta-learning enables the model to quickly update and adapt when faced with new tasks by learning "how to learn". In the present invention, the Model-Agnostic Meta-Learning (MAML) algorithm is used to initialize the model. The core idea of ​​MAML is to train a model so that its initial parameters are close to the "optimal adaptation point", that is, when the model faces a new task, only a small amount of gradient updates are required to quickly adapt to the new task. For example, after the model is trained on multiple different assembly tasks, its parameters will be in a "globally optimal" position. When a robot faces a new assembly task, the model trained based on MAML only needs to perform a small amount of local data training (such as a few iterations) to adapt to the requirements of the task.

[0118] The advantage of the meta-learning strategy is that after receiving the global model, when faced with new tasks or environmental changes, the model only needs a small amount of training data and a small number of training steps to quickly adjust to the new task. This feature is particularly suitable for dynamically changing industrial scenarios. For example, a robot originally performs a screw assembly task and then needs to switch to an electronic component welding task. Under the MAML strategy, the model can quickly adjust its end control strategy and force feedback adjustment parameters to adapt to different task requirements without having to train a new model from scratch. This significantly improves the robot's task switching efficiency and adaptability.

[0119] S440, based on a continuous learning mechanism, dynamically adjusting the adaptively trained robot to obtain a robot with multi-task compatibility;

[0120] Specifically, S441, based on the continuous learning mechanism, the learning rate of the robot after adaptive training is dynamically adjusted to obtain the adjusted robot; S442, the knowledge distillation algorithm is used to perform knowledge transfer learning between models on the adjusted robot to obtain a robot with multi-task compatibility.

[0121] In step S440 of some embodiments, for example, in order to ensure that the robot always has the latest knowledge and optimal model performance in a dynamic environment, the system has designed a continuous learning mechanism. During the execution of the task, the robot can continuously learn and update its model to avoid the "model forgetting" phenomenon.

[0122] During task execution, the robot continuously collects new sensor data, assembly results, and operational feedback, using this new data to fine-tune the model online. Through this continuous learning mechanism, the model can continuously evolve to adapt to changes in the environment and tasks. To prevent overfitting or forgetting during model updates, online updates typically employ incremental learning strategies to ensure that the model retains existing knowledge when new data is introduced. For example, during welding, if the robot encounters new assembly materials or environmental changes, sensor feedback such as temperature and force will be updated in real time to the model. The model then adjusts the welding force or time based on this new data to ensure optimal welding quality.

[0123] During online updates, the system dynamically adjusts the learning rate to ensure model stability and robustness. For significant new changes (such as changes in material properties or assembly tool switching), the system increases the learning rate to accelerate adaptation; for minor environmental fluctuations, the system decreases the learning rate to avoid drastic changes in the model.

[0124] During the continuous updating process, the system introduces knowledge distillation technology to prevent the model from forgetting knowledge from old tasks when encountering new tasks. The core idea of ​​knowledge distillation is to utilize the knowledge transfer between the new and old models, retaining important knowledge from the old model through soft labels, and preventing model degradation.

[0125] Specifically, when the model is trained on new task data, the system uses the output of the old model as an auxiliary target, allowing the new model to simultaneously adapt to the new task and maintain the performance of the old task. This method can effectively solve the model's "catastrophic forgetting" problem, ensuring that the robot always has comprehensive task execution capabilities when faced with diverse tasks. For example, the robot initially focused on precision parts assembly. When it receives a completely new type of assembly task, knowledge distillation ensures that learning the new task does not affect the robot's mastery of the precision parts assembly task, allowing it to switch smoothly between new and old tasks.

[0126] Through knowledge distillation, the model can retain important features across multiple tasks, which is particularly important for robots that need to perform multiple assembly tasks simultaneously. The system enables the robot to quickly switch between tasks while performing one task and retain key knowledge for other tasks, thereby improving task compatibility and responsiveness.

[0127] S450: Execute assembly tasks based on a robot with multi-task compatibility to obtain robot assembly optimization results.

[0128] In addition, it is also necessary to point out that in the model updating process, noise is added to protect the privacy of local data, strict permission management is further set, only authorized entities can access the model and data, and the system state is monitored in real time to discover faults and alarm in time. The hot backup and automatic switching of the support system are supported to ensure continuous operation, and finally the fault tolerance mechanism is set. In the process of federated learning, if a robot is offline or fails, the system can still normally perform model aggregation and updating.

[0129] In summary, the embodiments of the present application have the following improvements over the prior art:

[0130] 1) High-resolution cameras, laser radars, and force sensors are used to realize high-precision real-time perception of the assembly environment. The camera is used to obtain visual information of the assembly object, the laser radar is used for three-dimensional environment modeling, and the force sensor is used to monitor the force feedback of the end effector, ensuring that the robot has comprehensive perception ability for position, angle, and force changes during assembly. The time synchronization, spatial alignment, and data fusion algorithm of multi-sensor data are processed through a deep learning model to improve the accuracy and robustness of environment perception.

[0131] 2) In the multi-robot cooperation framework, federated learning is used to realize distributed model training and sharing of each robot. Each robot trains a multi-sensor fusion model and a motion control model locally, encrypts and uploads the updated model parameters to the central server, performs weighted aggregation through the asynchronous federated averaging algorithm, generates a global model and distributes it to each robot, realizes knowledge sharing, adopts the asynchronous federated averaging (FedAsync) strategy, and weighted aggregates the model parameters uploaded by each robot, combines data encryption and compression transmission technology to ensure data security and model privacy. In the aggregation process, the Byzantine defense algorithm (such as Krum algorithm) is applied to filter abnormal data to ensure the stability and security of the global model.

[0132] 3) The personalized model adopts a personalized structure design of shared layers and special layers. The first few layers are shared layers that extract general features, and the last few layers are special layers that are optimized for the individual needs of robots in specific assembly tasks and environments, making the model both general and task-adaptive. The shared layer updates use the global model optimization of federated learning, and the special layer is fine-tuned by local data to enhance the response ability and adaptability of the robot to specific tasks. The personalized model structure realizes task customization of individual robots without affecting the sharing of knowledge in federated learning.

[0133] 4) During the assembly task execution, real-time feedback from force sensors and vision sensors is used to detect assembly errors and automatically compensate. The force sensor monitors the force changes during assembly, and the vision sensor monitors the position and attitude of the assembly object. The system automatically adjusts the motion trajectory of the robot arm based on the error calculation results through the PID controller or Model Predictive Control (MPC) algorithm. The error detection module combines force and vision information to calculate the deviation vector in real time. The compensation algorithm quickly corrects the motion path through the controller to ensure high precision operation of the robot during assembly. The error compensation module predicts historical error patterns through deep learning to adjust in advance, improving the speed and accuracy of compensation.

[0134] 5) Based on the Model-Agnostic Meta-Learning (MAML) meta-learning algorithm, the robot model is trained to have rapid adaptation capability, so that it can quickly adjust to new task requirements through a small number of training steps when the task switches or the environment changes. The meta-learning algorithm plays a role in the global model initialization process, so that the model parameters are close to the "optimal adaptation point", ensuring that the model can quickly update and adapt when facing new tasks. The meta-learning strategy effectively improves the response speed and adaptability of the system in the multi-task switching scenario.

[0135] 6) During the execution of the assembly task, the robot continuously collects new data for incremental training through online learning to prevent the "model forgetting" phenomenon. At the same time, knowledge distillation technology is used to realize knowledge transfer between old and new models to prevent the model from forgetting old task knowledge during new task learning. The robot system applies knowledge distillation when updating the model, using the soft labels (Soft Labels) of the old model as auxiliary targets for the new model to realize compatibility between new and old tasks, ensuring that the model always maintains stable performance in multi-task execution.

[0136] 7) In the multi-robot collaboration scenario, the central server intelligently analyzes and allocates the task state. The system schedules and allocates tasks based on the task load, assembly success rate, and real-time feedback of each robot to ensure the balance of tasks and overall assembly efficiency. The task allocation system dynamically adjusts the task allocation strategy based on real-time state information, historical data, and environmental feedback of the robot. Through real-time information sharing, robots can actively avoid collisions and dynamically adjust actions during collaboration to ensure efficient and safe collaborative work.

[0137] 8) The robot system uses the AES encryption algorithm in data transmission to ensure the security of model parameters during uploading and issuing, and combines the SSL / TLS protocol to establish an encrypted communication channel, avoiding data leakage and attacks, and encrypting the transmission of model parameters to prevent unauthorized access and data tampering. Through the secure communication protocol, the privacy of the federated learning model update is protected, and the data security and system integrity of the multi-robot system are ensured.

[0138] Please refer to Figure 2 The embodiment of the application also provides a robot assembly optimization system based on federated learning, which can implement the above-mentioned robot assembly optimization method based on federated learning. The system comprises:

[0139] The first module 201 is configured to acquire image data and point cloud data of the robot and perform data feature fusion and recognition processing to obtain comprehensive feature data of the robot.

[0140] The second module 202 is configured to acquire force sensation information of the robot and visual information of the robot, and perform motion error optimization processing based on deep reinforcement learning in combination with the comprehensive feature data of the robot to obtain comprehensive feature data of the robot after error compensation.

[0141] The third module 203 is configured to perform model training on the robot based on the comprehensive feature data of the robot after error compensation through a federated learning optimization algorithm to construct a robot with a global assembly control model.

[0142] The fourth module 204 is configured to perform an assembly task in communication and cooperation based on the robot with the global assembly control model based on a meta-learning strategy to obtain a robot assembly optimization result.

[0143] It can be understood that the contents in the above method embodiments are all applicable to the system embodiments, the system embodiments specifically implement the same functions as the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0144] The preferred embodiments of the application are described above with reference to the accompanying drawings, but this does not limit the scope of the application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the application shall be within the scope of the application.

Claims

1. A robot assembly optimization method based on federated learning, characterized in that: The method comprises the following steps: Acquire the robot's image data and point cloud data and perform data feature fusion and recognition processing to obtain the robot's comprehensive feature data; Acquiring force information and visual information of the robot, and combining the robot's comprehensive feature data with motion error optimization processing based on deep reinforcement learning to obtain error-compensated robot comprehensive feature data; Based on the comprehensive feature data of the robot after error compensation, the robot model is trained through a federated learning optimization algorithm to construct a robot with a global assembly control model; Based on the meta-learning strategy, the robots with the global assembly control model communicate and collaborate to perform assembly tasks, thereby obtaining robot assembly optimization results; The acquisition of the robot's force information and the robot's visual information, and combining the robot's comprehensive feature data with motion error optimization processing based on deep reinforcement learning to obtain error-compensated robot comprehensive feature data, include: Acquiring force information of the robot through a force sensor and acquiring visual information of the robot through a visual sensor; Defining a state space for deep reinforcement learning based on the comprehensive feature data of the robot, defining an action space for deep reinforcement learning based on the force information and visual information of the robot, setting an experience replay pool, and conducting interactive learning to obtain preliminarily optimized comprehensive feature data of the robot; Controlling the robot to perform simulation according to the preliminary optimized comprehensive characteristic data of the robot to obtain a robot motion deviation vector; The robot motion deviation vector is error compensated by an error compensation algorithm, and is optimized in combination with a deep reinforcement learning algorithm to obtain the comprehensive feature data of the robot after error compensation.

2. The method according to claim 1, characterized in that The acquisition of the robot's image data and point cloud data and the data feature fusion and recognition processing to obtain the robot's comprehensive feature data includes: Calibrate the camera and lidar to obtain a calibrated camera and a calibrated lidar; Acquiring image data of the robot through the calibrated camera and acquiring point cloud data of the robot through the calibrated lidar; Performing data preprocessing and data fusion processing on the image data of the robot and the point cloud data of the robot to obtain fused robot data; Introducing image branch module, point cloud branch module and feature fusion layer to build a deep learning fusion model; Based on the deep learning fusion model, feature recognition and fusion processing are performed on the fused robot data to obtain the comprehensive feature data of the robot.

3. The method according to claim 1, characterized in that The method includes: performing model training on the robot through a federated learning optimization algorithm based on the comprehensive feature data of the robot after error compensation to construct a robot with a global assembly control model, including: Controlling the robot to perform an assembly task according to the error-compensated comprehensive feature data of the robot, and constructing a local data set according to the assembly execution result; Performing data preprocessing on the local data set to obtain a preprocessed local data set; A robot assembly control model is constructed based on the multi-sensor fusion model and the motion control model, and trained using a small-batch gradient descent algorithm to obtain the trained robot assembly control model. The pre-processed local data set is compressed and uploaded to the trained robot assembly control model through the AES symmetric encryption algorithm; Based on the trained robot assembly control model, assigning weight parameters to the preprocessed local data set and performing aggregation processing through an asynchronous federated averaging algorithm to obtain a global assembly control model; The global assembly control model is sent to a robot control terminal to obtain the robot having the global assembly control model.

4. The method according to claim 1, wherein The meta-learning strategy is used to communicate and collaborate with the robots having the global assembly control model to perform assembly tasks, thereby obtaining robot assembly optimization results, including: Based on an industrial communication protocol, the robots having the global assembly control model are networked for communication; Introducing a personalized network model, performing global network fine-tuning processing on the robot having the global assembly control model, and obtaining a fine-tuned robot; A meta-learning strategy is introduced to perform adaptive training on the fine-tuned robot to obtain an adaptively trained robot; Based on a continuous learning mechanism, dynamically adjusting the adaptively trained robot to obtain a robot with multi-task compatibility; The robot assembly optimization result is obtained by executing the assembly task based on the robot with multi-task compatibility.

5. The method according to claim 4, characterized in that The personalized network model includes a plurality of shared layers and a plurality of dedicated layers, wherein the output ends of the plurality of shared layers are connected to the input ends of the plurality of dedicated layers, wherein: The shared layers are used to extract common features of the robot; The dedicated layers are used to extract the personalized demand features of the robot.

6. The method according to claim 4, characterized in that The introducing of the meta-learning strategy and the adaptive training of the fine-tuned robot to obtain the adaptively trained robot include: Initializing the fine-tuned robot using a meta-learning algorithm to obtain an initialized robot; The end control and force feedback parameters of the initialized robot are adjusted to obtain the adaptively trained robot.

7. The method according to claim 4, characterized in that The method of dynamically adjusting the adaptively trained robot based on the continuous learning mechanism to obtain a robot with multi-task compatibility includes: Based on a continuous learning mechanism, dynamically adjusting the learning rate of the adaptively trained robot to obtain an adjusted robot; The adjusted robot is subjected to inter-model knowledge transfer learning through a knowledge distillation algorithm to obtain the robot with multi-task compatibility.

8. The method according to claim 1, characterized in that The method also includes adding noise during the process of the robots with the global assembly control model communicating and cooperating to perform assembly tasks.

9. A robot assembly optimization system based on federated learning, characterized in that: The system comprises: The first module is used to obtain the robot's image data and point cloud data and perform data feature fusion and recognition processing to obtain the robot's comprehensive feature data; The second module is used to obtain the robot's force information and the robot's visual information, and combine the robot's comprehensive feature data to perform motion error optimization processing based on deep reinforcement learning to obtain the robot's comprehensive feature data after error compensation; A third module is used to perform model training on the robot through a federated learning optimization algorithm based on the comprehensive feature data of the robot after error compensation, so as to construct a robot with a global assembly control model; The fourth module is used to communicate and cooperate with the robots with the global assembly control model to perform assembly tasks based on a meta-learning strategy, thereby obtaining robot assembly optimization results. The acquisition of the robot's force information and the robot's visual information, and combining the robot's comprehensive feature data with motion error optimization processing based on deep reinforcement learning to obtain error-compensated robot comprehensive feature data, include: Acquiring force information of the robot through a force sensor and acquiring visual information of the robot through a visual sensor; Defining a state space for deep reinforcement learning based on the comprehensive feature data of the robot, defining an action space for deep reinforcement learning based on the force information and visual information of the robot, setting an experience replay pool, and conducting interactive learning to obtain preliminarily optimized comprehensive feature data of the robot; Controlling the robot to perform simulation according to the preliminary optimized comprehensive characteristic data of the robot to obtain a robot motion deviation vector; The robot motion deviation vector is error compensated by an error compensation algorithm, and is optimized in combination with a deep reinforcement learning algorithm to obtain the comprehensive feature data of the robot after error compensation.

Images