The key to complex product assembly based on digital twin is to construct a cyber-physical fusion system, whose assembly process is a physical and virtual assembly co-evolutionary process. In order to achieve "controllable and measurable" in assembly quality, it is first necessary to build a multi-dimensional, multi-scale, and multidisciplinary high-fidelity model of complex product, and use the digital twin high-fidelity model for assembly simulation; Secondly, it is necessary to realize the real-time collection, sensing and fusion of multi-source heterogeneous data in the assembly process; Finally it is necessary to combine physical and virtual data fusion technology to realize dynamic data analysis, decision-making and optimization, and build an online accurate prediction model of assembly performance, realize intelligent simulation and online accurate prediction of assembly performance based on digital twin.
4.1 Multidimensional, multi-scale, multidisciplinary modeling and simulating of digital twin-driven assembly
The core of the digital twin is the mapping of physical entities in virtual space. The fidelity of the virtual model determines the success of the application of digital twin assembly technology. In the digital twin-driven assembly process, the virtual assembly workshop guides the physical assembly workshop to complete assembly tasks by simulating the assembly behavior of the physical assembly workshop. Therefore, establishing a high-fidelity virtual model of the virtual assembly workshop based on geometry, physics, behavior, and rules, and satisfying the simultaneous evolution of the virtual model and physical perception data are the core of realizing the quality control and prediction of complex product assembly based on the digital twin.
Unlike traditional complex products, the assembly problems of complex product are not only reflected in pure structural performance such as structural fatigue, strength, and stiffness, but also in electrical performance such as electromagnetic and signal processing, and electromechanical coupling. Therefore, in complex product assembly quality control and prediction, it is necessary to establish a multi-scale (system, sub-assembly, parts, etc.), multidimensional (function, performance, behavior, etc.), multidisciplinary (mechanical, electromagnetic, etc.) virtual model. At the same time, the virtual model is required to have multi-scale, multidimensional, multidisciplinary simulation, analysis, and prediction capabilities. The simulation of the assembly process in the virtual environment needs to ensure mechanical performance, and needs to ensure or even improve electrical performance to realize the mechatronics of complex product assembly.
Each physical performance of the complex product assembly process corresponds to a specific model, such as assembly error model, assembly deformation model, electromagnetic performance model and electromechanical coupling model. It is the key to the quality prediction and control of complex product assembly that how to associate these models with different physical characteristics. The multi-physics integrated modeling can not only improve the accuracy of virtual simulation, and the simulation results can also reflect the state and behavior of the physical assembly more realistically, making it possible to replace the physical prototype with digital twin-driven assembly. Therefore, multi-scale, multidimensional, multidisciplinary modeling and simulation will be an important technical mean to improve the fidelity of virtual models and give full play to the role of digital twin.
4.2 Multi-source heterogeneous data collection, sensing and fusion for assembly processes
The collection and transmission of high-fidelity sensor data is the basis for establishing a digital twin model of complex product. The complex product assembly process generates a large amount of assembly data, which include soft sensor data such as simulation data, calculation data and mechanism data of the digital twin workshop, and actual measurement data of the physical assembly workshop. With the help of OPC communication protocol, RFID, ultrawide band (UWB) and other technologies[65, 66], the assembly process quality data will be collected and stored in the virtual assembly workshop and the physical assembly workshop, and through the analysis of each assembly quality data, the reduced-order characterization model between complex product parts will be constructed, and then the coupled reduced-order accuracy correction will be completed based on deep learning algorithms to establish the agent model in the assembly process of complex product, which is an important way to realize online quality control and prediction of complex product assembly process based on digital twin. In addition, in the process of complex product assembly, the data collected by various sensor devices has the characteristics of grammatical, semantic, and structural heterogeneity, which is a huge challenge for multi-source heterogeneous data collection and fusion.
The assembly process quality data collection methods mainly include software collection, hardware collection and manual collection. Software collection is the indirect collection from other systems through software integration interfaces and database sharing; hardware collection is the direct collection of assembly process quality data through measurement, sensors, and other instruments; manual collection refers to part of the assembly quality data needs to be manually recorded and stored in the actual assembly process, and manual collection mainly relies on manual measurement and input. The assembly process quality data collection process is shown in Fig. 2.
In the process of complex product assembly, the types, attributes, and content of collected data are very different. So, it is necessary to study the standard data communication and conversion interface, realize the unified conversion and unified encapsulation of different types of data, and convert the multi-source heterogeneous data into definite data that computer-recognizable and required for digital twin simulation analysis. On this basis, it is also necessary to plan, clean, excavate, cluster, and package the real-time data, simulation data and mechanism data covering the assembly process of complex product to achieve data traceability and operability. Through the operation of classification, association, and combination of multi-source heterogeneous data, it is helpful to accurately depicts and reflects the state of dynamic evolution process, evolution law, statistical features, and other development process, and improve the accuracy of state and feature estimation, and provide real-time basic data support for the construction of digital twin model.
4.3 Data-driven decision making, feedback and optimization technology
In the assembly process of complex product, the assembly progress, assembly accuracy and reliability are monitored, adjusted, and optimized based on theoretical values and mechanistic models through the interactive fusion, analysis and decision making of multi-source heterogeneous data. The digital twin-driven feedback technology assists assemblers to strengthen the closed-loop feedback control of the assembly process through the comparative analysis of physical data and design data, which will realize real-time monitoring of the assembly process and achieve quality control and online accurate prediction of the assembly performance.
During the assembly stage, various sensors are used to track and monitor the assembly status such as mechanical properties, electromagnetic properties, environmental conditions, and operating parameters. These monitoring data and historical data can predict analyze the assembly performance and other indicators. Based on the digital twin data that fuse by physical measurement data and virtual simulation data, the latest measurement data, schedule data, performance data, and measurement values of assembly process state parameters are mapped to the digital twin model, and this will realize a closely connection with the real world and the information world based on the digital twin assembly model. Similarly, the big data analysis methods such as Markov method, D-S evidence theory[68–70], Taguchi method[71, 72], support vector machine will also be used in the stage of data analysis and decision making to achieve assembly quality control and prediction, which helps reduce assembly errors and improve assembly efficiency.