In complex assembly industry settings, fault localization involves rapidly and accurately identifying the source of a fault and obtaining a troubleshooting solution based on fault symptoms. This study proposes a knowledge-enhanced joint model that incorporates aviation assembly knowledge graph (KG) embedding into large language models (LLMs). This model utilizes graph-structured Big Data within KGs to conduct prefix-tuning of the LLMs. The KGs for prefix-tuning enable an online reconfiguration of the LLMs, which avoids a massive computational load. Through the subgraph embedding learning process, the specialized knowledge of the joint model within the aviation assembly domain, especially in fault localization, is strengthened. In the context of aviation assembly functional testing, the joint model can generate knowledge subgraphs, fuse knowledge through retrieval augmentation, and ultimately provide knowledge-based reasoning responses. In practical industrial scenario experiments, the joint enhancement model demonstrates an accuracy of 98.5% for fault diagnosis and troubleshooting schemes.

In this article, a novel work pattern for large-scale workpiece machining, that is, multimobile robot cluster machining, is proposed. The key idea is to use distributed coordination control to link multiple robots as a cluster system. The coordination control law acts as a virtual constraint, and multiple robots move synchronously under a common leader. Two topological structures are introduced to describe the communication networks of the cluster system, namely, the N -chain and close-loop topologies. The convergence rate and resistance index of different topological structures are analyzed, and the applications of different topologies in varied machining scenarios are discussed. Two experiments are conducted on the multimobile robot system. The cluster system with a close-loop topology is used for the collaborative handling task, where the relative distance of two mobile robots can be limited within 6 cm. The cluster system with an N -chain topology is used for large-scale wind turbine blade machining. The cluster machining system requires little manual intervention and guarantees the synchronization performance of the machining system.

In this article, a mobile robotic grinding system for large-scale workpiece is proposed. The mobile robot is equipped with a laser sensor, fringe projection scanner, and force-controlled grinding tool. The mobile robot can move in large spaces, calibrate the workpiece accurately, and perform grinding work automatically. A vision-based grinding strategy for the mobile robot is proposed, where the workpiece surface reconstruction and grinding path planning are generated by the measured point clouts. An error index is proposed to analyze the machining accuracy, which can reveal the primary error factor in the error transfer chain. Compared with the mobile robotic grinding based on the global measurement system in (Zhao et al., 2021), the proposed vision-based grinding process requires no precise CAD model of the workpiece and achieves better machining accuracy. Two experiments are conducted in this article, where a wind turbine blade and a part of the aircraft skin have been ground by the mobile robot. Experiments show that the mobile robot under the vision-based grinding strategy is suited for different continues surfaces, and it has great application prospects.

In complex robot applications, such as human−robot interaction and robot machining, robots should interact with an unknown environment. To learn the interactive skill, a model-based actor−critic learning algorithm and a safety-learning strategy are proposed in this article to find the optimal impedance control, in which the learning process is safe and fully automatic and does not know the system parameter. In the learning algorithm, a critic is defined as a quadratic form of the system states and the external force. A modified deterministic policy gradient algorithm is presented to improve the learning efficiency. The proposed approach utilizes a model-based constraint and a highly efficient learning algorithm. In the safety-learning strategy, the robot is trained under a constant force, and the learned impedance control can transfer to different interaction situations by choosing the suitable impedance index. The effectiveness of the learning algorithm and the performance of the learned impedance control are validated in a UR5 robot. The robot can perform human−robot interaction and robot machining tasks after the training process with 100 s training time.

Dual-arm robots show better task adaptability but face more constraints than single-arm robots. Inspired by human cooperation options, a nonlinear model predictive cooperative control (NMPCC) coupled with a cooperative index is proposed in this article. By adjusting the cooperative index, dual-arm cooperation can be classified into four modes: tracking, motion synchronization, impedance, and force priority. Thus, the cooperation operation problem is converted to a multiobjective optimization problem. Then, a nonlinear model predictive solver coupled with the ACADO toolkit is designed to solve the multiobjective optimization problem, where the robotic torque control input can be calculated in real time (less than 1 ms). The motion synchronization, cooperative transportation, and human dual-arm robot interaction experiments were conducted on dual Franka panda robots. Experiments reveal that the NMPCC control coupled with the cooperative index is easy to apply in dual-arm robots and can adapt to different complex manipulation tasks.

In robotic manufacturing, the machining vibration can easily affect the quality of the processed surfaces due to the poor stiffness and variable stiffness of the robot arm. It is important to monitor the processing quality. Therefore, an online process monitor method for robotic grinding is proposed in this paper. Since the surface quality is difficult to be measured directly online, a vibration-surface quality map is proposed to show the mapping relation between the vibration signals and the surface images. Based on this map, the grinding states are classified to lack, normal and over grinding, which formed the training set data. Then a Convolutional Neural Network (CNN) classifier is trained to detect the grinding states during robotic machining only with the vibration signals. To establish the vibration-surface quality map, the Symmetrized Dot Pattern (SDP) and Sparse Auto-Encoder (SAE) methods are utilized to extract the features from the raw vibrations and images data, and a Gray-Level Co-occurrence Matrix (GLCM) index is further used to build the bridge between them. The experimental results in this paper have proved the effectiveness and potential of our method for online robotic grinding quality monitoring.