1. Introduction With the development of advanced manufacturing technology, the automation, flexibility and intelligence of welding product manufacturing has become an inevitable trend [1-8]. At present, the use of robotic welding has become the main symbol of the modernization of welding automation technology. Welding robots have received increasing attention due to their advantages of strong versatility and reliable operation. The use of robotic technology in welding production can improve productivity, improve working conditions, stabilize and ensure welding quality, and realize the automation of welding of small batch products [9]. From its birth and development in the 1960s to the present, the research on welding robots has gone through three stages, namely the teaching and playback stage, the offline programming stage and the autonomous programming stage. With the continuous progress of computer control technology, welding robots are developing from single-machine teaching and playback type to multi-sensor, intelligent flexible processing units (systems), and the transition from the second generation to the third generation will become the goal pursued by welding robots [9,10]. Currently, most arc welding robot systems used domestically and internationally are first-generation or near-second-generation welding systems. Since the welding path and parameters are pre-set based on actual operating conditions, they lack external information sensing and real-time adjustment and control functions during welding. These arc welding robots have strict requirements for the stability of welding operating conditions and lack "flexibility" during welding, exhibiting obvious shortcomings. In actual arc welding, welding conditions frequently change. For example, errors in processing and assembly can cause changes in the position and size of the weld, and changes in the workpiece's heating and cooling conditions during welding can cause weld deformation and uneven penetration [9,12]. To overcome the impact of various uncertainties on welding quality during robot welding and improve the intelligence level and reliability of robot operations, arc welding robot systems must not only achieve automatic real-time tracking of spatial welds but also realize online adjustment of welding parameters and real-time control of weld quality. 2. Main Components of Intelligent Robot Welding Technology Modern welding technology is characterized by typical multidisciplinary integration [5,11], and the adoption of robot welding is a concentrated manifestation of the technological achievements of related disciplines. The main technological components involved in introducing intelligent technology into welding robots are shown in Figure 1. These include: 1) Autonomous planning technology for welding tasks by welding robots; 2) Motion trajectory control technology for welding robots; 3) Information sensing, modeling and intelligent control technology for welding dynamic processes; 4) Integration and control of robot welding systems, integrating the hardware and software design, unified optimization scheduling and control of the above-mentioned welding task planning, trajectory tracking control, sensing system, process model, intelligent control and other subsystems, involving the management and control of material flow and information flow in flexible welding manufacturing systems, and the control of multiple intelligent units and complex systems with multiple robots, sensors and controllers. The main research and development status of the above-mentioned related technologies are briefly described below. 3. Autonomous planning technology for robot welding tasks [13-16] As mentioned above, most arc welding robots at home and abroad are of the teach-and-reproduce type, which cannot meet the increasingly complex needs of welding production, and there are still many problems to be studied. Research on arc welding robots has gradually transitioned towards autonomy, leading to the emergence of offline programming technology for arc welding robots. A relatively complete offline programming system for arc welding robots should consist of several major parts, including welding task description (language programming or graphical simulation), operator-level path planning, kinematic and dynamic algorithms and optimization, joint-level planning for the welding task, animation simulation of planning results, offline correction of planning results, communication interface with the robot (downloading), autonomous path planning using sensors, and online path correction. Key technologies typically include the design of vision sensors and the acquisition of weld seam information, as well as the design of the planning controller. At the 1987 International Conference on Automation and Robotic Welding, experts summarized the development of offline programming, with the WRAPS system being the most representative work. KHGoh et al. established an expert system-based adaptive offline programming and control system for welding robots—WRAPS—on a FANUC/WESTWOOD welding workstation. It included a welding database, offline programming, computer simulation, and a welding expert system. It was also equipped with vision sensors for pre-weld joint inspection and post-weld defect detection, thus forming a complete expert welding robot system. Researchers at Harbin Institute of Technology in China have conducted research on collision-free path planning for welding robots, autonomous planning of redundant arc welding robots, and joint planning of welding process parameters [14-16]. They designed and developed an offline planning and simulation system module structure, as shown in Figure 2. Figure 2: Structure of the offline planning and simulation system. Arc welding robot planning systems typically include a CAD input system, a welding expert system, an autonomous planning system, and a simulation system. More broadly, a more comprehensive arc welding robot planning system should also include a feedback control system, a pre-weld sensing system, and a post-weld inspection system. 4. Weld Seam Tracking and Guiding Technology in Robot Welding. In terms of robot welding operations, the motion trajectory control of welding robots mainly refers to initial weld position guidance and weld seam tracking control technology. In various application areas of arc welding robots, adaptability is the most important factor affecting welding quality and efficiency. The adaptability of an arc welding robot refers to the real-time control and correction of the robot's operation using input signals from sensors detected from the welded workpiece, to adapt to changing welding conditions and environments. Swedish and American companies have successively developed laser scanning and structured light vision sensors for weld seam tracking systems. Dr. Lü Weixin and Dr. Zhang Jiong from Harbin Institute of Technology developed a vision system based on laser scanning and a high-performance linear array CCD sensor, as shown in Figure 4, to achieve real-time visual control [13]. In the research on the initial welding position robot visual guidance technology, Dr. Lü Weixin designed a local search algorithm based on the laser scanning vision system to achieve autonomous guidance of the weld seam features of a certain workpiece within a certain range [13]. Guo Zhenmin and Li Jinquan respectively used visual servo and image recognition technology to explore the problems of robot welding initial welding position guidance and weld seam recognition and real-time tracking. 5. Visual sensing, modeling and intelligent control technology of welding pool dynamic process The key to high quality of robot welding lies in achieving effective and accurate control of the welding dynamic process. Due to the complexity of the welding process, practice shows that the effectiveness of classical control methods is greatly limited. Inspired by the operating skills of skilled welders, in recent years, intelligent control methods that simulate welder operation have been introduced into the welding dynamic process, mainly involving visual sensing, modeling and intelligent control of the welding pool dynamic process. 5.1 Sensing technology of welding process Sensing of the welding process is a key link in achieving welding process quality control. The future trend of welding development is welding automation, robotics and intelligence, and sensing technology is the most important part of this development. Welding sensors can be divided into two categories according to their purpose: measuring and detecting the operating environment and detecting and monitoring the welding process. In terms of sensing principles, they are mainly divided into acoustic, mechanical, arc and optical sensing. Acoustic sensors are mainly used for detecting droplet transfer in GMAW process, plasma perforation welding, etc. Mechanical sensing mainly refers to the molten pool oscillation method that has been developed in recent years. Arc sensors directly detect the characteristics of the arc itself (current, arc voltage), without the need for external sensors and the protection and noise reduction devices required for external sensors, making the application simple. The current application areas are mainly weld seam tracking and deposition control. Compared with other sensing methods, optical sensors do not contact the welding circuit, and the detection of signals does not affect the normal welding process. It is one of the most promising sensing technologies in the future. Using infrared thermal radiation of the welding zone to sense welding process information is a self-contained method in optical sensing, and there are many research results in this area [16]. 5.2 Visual Sensing of Welded Molten Pool For the welding process, direct vision is the best non-contact sensing method. The main advantages of direct vision sensing technology are that it does not contact the workpiece, does not interfere with the normal welding process, acquires a large amount of information, and has strong versatility. Furthermore, because it can obtain two-dimensional or three-dimensional information of the dynamic molten pool during the welding process, compared with other welding process information detection methods, the molten pool information detected by this method directly reflects the dynamic behavior of the molten metal during the welding process, making it more suitable for quality control of the welding process. Applications of direct vision sensing in welding include offline determination of the position of the workpiece to be welded; online compensation for welding path deviations caused by fixed accuracy, tolerances of various parts of the robot, and deformation of the workpiece during the welding process; real-time sensing of the geometry of the weld joint and molten pool in welding process control; and monitoring of droplet transfer patterns. In recent years, with the development of computer vision technology, using machine vision to directly observe the welded molten pool from the front and obtaining the geometric information of the molten pool through image processing for closed-loop control of welding quality has become an important research direction [17-21]. Based on whether the imaging light source in the visual inspection system is an auxiliary light source or a light source generated by the welding area itself, direct vision inspection systems can be divided into two main categories: active and passive. (1) Active direct vision sensing In order to reduce the impact of arc light on image quality, the active direct vision detection method uses auxiliary light sources such as lasers to artificially illuminate the welding area to improve image quality. Since lasers have the characteristics of single wavelength, good directionality and good coherence, using lasers as auxiliary light sources can obtain clearer images. Active vision is often limited in its application and promotion due to the high cost of the required equipment and the complexity of the system. It will not be described in detail here. See the references in [14, 15]. (2) Passive direct vision sensing As a practical technology for visual image sensing in the welding process, most research focuses on using the arc light itself to illuminate the welding area instead of adding an auxiliary light source, i.e., passive direct vision detection method. The research on passive direct vision sensing began in the mid-1980s. Researchers at home and abroad have directly used the workpiece gap in front of the molten pool to obtain the weld information of the welding area, and realized the weld tracking in the welding process according to the flashing of the arc light intensity at different distances in front of the molten pool; the microcomputer control system with CCD camera was used to observe and control the behavior of pulsed MIG/MAG welding molten pool; the University of Ohio in the United States developed an integrated vision sensing system placed inside the welding torch and coaxial with the electrode to observe the welding molten pool, and conducted preliminary research on TIG welding molten pool observation and MIG welding weld tracking. Reference [19] designed a simultaneous sensing system for visual images of the front and back of the molten pool (Figure 3), obtained clear images of the front and back of the pulsed GTAW molten pool (Figure 4), and conducted in-depth research on the extraction of two-dimensional feature size of the image. Reference [21] obtained the pulsed GTAW filler wire molten pool image (Figure 5) based on the above system. And conducted research on the processing algorithm for three-dimensional shape feature recovery of the image. The results of references [19-21] can be considered as a relatively systematic and successful application of computer vision sensing technology in the dynamic process of welding molten pool in recent times. 5.3 Modeling of the dynamic process of welding molten pool Due to the nonlinearity, uncertainty, time-varying nature and strong coupling of the dynamic process of welding molten pool, the mathematical model established by the traditional process modeling method cannot be used as an effective controllable model. This is also the main reason why welding process control has been a major problem that has plagued the welding and control fields for a long time. At present, it is believed that the modeling methods for the dynamic process of welding molten pool are generally as follows: (1) Analytical mathematical model based on the theory of heat conduction of welding molten pool metal. This model is generally described by a set of partial differential equations, which deviates greatly from the actual process. The main problem is that it is difficult to use for the design of welding process control system. It is generally only used for numerical simulation and analysis of welding thermal process; (2) Mathematical model obtained by system identification method of classical or modern control theory, such as transfer function, difference equation and other forms. Generally, it can be used for real-time control of welding processes within a limited range; (3) The artificial neural network modeling method based on the neural network approximation theory obtains the neural network model of the welding process. This modeling method requires less knowledge of the process. Usually, the network model can be learned based only on the input and output data of the process. The obtained model is convenient for the online learning and control of the system; (4) The knowledge modeling method based on fuzzy set and rough set theory can be used with the help of welder experience or directly based on the processing of experimental data measurement to extract knowledge rules and give a model described in the form of human knowledge. This helps to understand the change law of the welding process and the application of intelligent systems. Reference [22] identified and analyzed the traditional mathematical model of the dynamic process of pulse GTAW molten pool. Reference [20] studied the fuzzy logic and neural network modeling method of the dynamic process of pulse GTAW molten pool, gave the corresponding model as shown in the figure, and verified the effectiveness of the model for real-time process control. 1) Extraction of fuzzy control rules for pulse GTAW docking process [20]: Fuzzy control rules are the basis for the design of fuzzy control systems. The extraction of fuzzy control rules for the controlled object belongs to the problem of fuzzy system identification. The C-means dynamic clustering algorithm is used to extract the fuzzy control rules for pulse GTAW docking process. The controlled variable of the pulse GTAW docking process is selected as the maximum width of the back side of the molten pool, and the control quantity is the pulse duty cycle. According to the results of the process dynamic experiment [20], the fuzzy control rules are extracted by the C-means dynamic clustering method as shown in Table 1: 2) The neural network model BNNM for the back side width of the pulse GTAW molten pool [19], its input vector includes the maximum width Wfmax(t) and the maximum half length Lfmax(t) of the front side of the molten pool at the current time, as well as 17 input parameters such as process parameters. The output of the model is the maximum width Wbmax(t) of the back side of the molten pool. The number of hidden layer processing units of the BNNM model is determined to be 24. The obtained BNNM model structure is shown in Figure 6. Figure 7 is a comparison curve between the expected output value of the training samples and the output value of the BNNM model. The average relative error of the model output is 4.25%, and the mean square error is 3.04%, which meets the accuracy requirements. Figure 7 BNNM model verification results 5.4 Intelligent control of welding dynamic process Since the welding process is a time-varying nonlinear system with multiple parameters coupled together, there are many factors affecting the weld formation quality, and it has significant randomness, making it difficult to describe with an accurate mathematical model. This makes some previous linear control methods have shortcomings such as poor adaptability and high dependence on experience to varying degrees. Therefore, introducing intelligent control methods into welding process control is a very suitable approach. There are already methods using expert systems, fuzzy control, and neural network control. Expert systems utilize the professional knowledge and experience of the controlled object domain and use the knowledge representation and reasoning techniques of artificial intelligence expert systems to derive control actions. Fuzzy control is an early form of intelligent control. It draws on the fuzzy nature of human thinking and uses tools such as membership functions, fuzzy relations, fuzzy reasoning, and decision-making in fuzzy mathematics to derive control actions. Neural network control studies and utilizes certain structures and mechanisms of the human brain, along with human knowledge and experience, for system control. It represents the application of neural networks as a form of artificial intelligence in the control field. With the increasing self-organizing, mapping, and decision-making capabilities of pattern recognition, neural networks have shown significant potential advantages in intelligent control design and implementation. Since the mid-1980s, research on the application of fuzzy control in welding has been conducted abroad. Researchers have applied self-organizing fuzzy control methods to study MIG weld seam tracking; used visual sensors to sense molten pool information and employed fuzzy control methods to control the weld width in pulsed MIG welding; used fuzzy control for parameter planning of arc welding robots; and conducted research on fuzzy control of arc voltage parameters in pulsed gas metal arc welding and fuzzy control of CO2 welding process parameters. Artificial neural network (ANN) control studies and utilizes certain structures and mechanisms of the human brain, along with human knowledge and experience, for system control. Because control systems designed using neural networks exhibit good adaptability and robustness, they can handle complex control problems in high-dimensional, nonlinear, highly disturbed, uncertain, and difficult-to-model processes such as welding. Therefore, using neural networks to establish welding process models in welding quality control can solve problems that linear control methods cannot overcome. This model differs from previous fixed-structure mathematical models, making no assumptions about the welding process, thus reflecting the system characteristics more realistically. ANNs have achieved some success in the welding field. For example, GTAW process experiments on low-carbon steel showed good agreement between experimental and expected values. A TIG weld pool size estimation system was established using ANNs, and the relationship between the surface temperature of the welding zone and the weld pool size was derived from its designed neural network estimator, enabling simultaneous prediction of the front weld width, back weld width, and weld depth during the welding process. Furthermore, neural networks have also been applied in weld seam tracking and weld zone image processing. Since the 1990s, research in this area in the domestic welding industry has gradually emerged. Research has been conducted on welding quality control based on ANN technology. Static and dynamic models of the welding process have been established for GTAW process, realizing intelligent control of the front weld width under the condition of rapid change of heat dissipation conditions in GTAW welding process. References [19-20] designed a single neuron self-learning control system based on the learning and adaptation characteristics of neural networks, realizing effective control of the dynamic characteristics of pulsed GTAW surfacing weld pool; designed a multivariable intelligent control system combining neural network self-learning and expert system (Figure 8), realizing effective control of the dynamic changes of the front and back weld width of pulsed GTAW butt weld pool; Reference [21] designed a fuzzy neural network adaptive intelligent control system (Figure 9), realizing effective control of the front and back weld width and front weld reinforcement height of pulsed GTAW butt filler wire weld pool; designed a fuzzy neural network adaptive intelligent control system, realizing effective control of the front and back weld width and front weld reinforcement height of pulsed GTAW butt filler wire weld pool; 6. Intelligent Robotic Welding Flexible Manufacturing Unit/System The main purpose of this research on the system integration and optimization technology of arc welding robot flexible processing units is to combine real-time welding quality control with robotics technology. It aims to study and implement an arc welding robot motion control system with redundant degrees of freedom and corresponding ship-shaped welding posture coordination control technology. Simultaneously, it integrates subsystems such as high-precision laser scanning weld seam tracking, intelligent control of penetration and weld seam formation, and robot welding power supply through a network into an arc welding robot flexible processing unit with real-time sensing, communication, and scheduling functions. The research also studies the optimization model and control strategy of a multi-variety, small-batch flexible welding processing system under the control of a central monitoring computer, achieving intelligent quality control of robotic welding of spatial curve weld seams. Based on the characteristics of welding environment and process sensor information acquisition, feature extraction, decision control, and process implementation, a three-level interactive hierarchical structure for the intelligent robotic welding technology system can be established—namely, the organization level, coordination level, and execution level. Building upon the research on program control and communication based on a real-time expert system environment in flexible manufacturing systems, this research studies knowledge-based control of arc welding robot systems, proposing four modules: modeling, welding planning, program generation, and communication. Generally, according to the different forms of continuous or discrete material flow (or energy flow) in the production process, manufacturing systems can be divided into three categories: continuous, discrete, and hybrid. Welding Flexible Manufacturing Cell/System has discreteness on a macroscopic level and continuity on a microscopic level. Since the control of the microscopic continuous welding process has achieved a lot, it is of great significance to study the discrete production process of large systems from the perspective of adapting to small batches and multiple varieties of welding products, in order to improve the utilization rate of welding flexible manufacturing systems and improve the quality of welding products. Discrete event dynamic systems and Petri net (PN) theory modeling and control have become an important research direction of modern manufacturing system design theory [25]. References [23-24] combined the discrete control theory of flexible manufacturing systems to conduct a systematic study on the construction, integration and real-time scheduling control technology of welding flexible processing units with multi-sensor information. A complete nine-degree-of-freedom arc welding robot motion control system was designed and developed. Under the guidance of the open structure system design concept, the spatial weld coordination control and the integrated scheduling of welding flexible processing units (WFMC) were realized. This study combines workpiece initial positioning spatial guidance technology, laser scanning weld real-time tracking technology, penetration and weld formation intelligent control technology, nine-degree-of-freedom arc welding robot system control technology, and welding robot dedicated power supply technology to investigate the integration method and communication control implementation of the welding flexible processing system. A complete welding flexible processing unit system with multi-sensor control functions and a corresponding central monitoring software platform were established (Figure 10). Figure 10 Hierarchical structure of intelligent arc welding robot WFMC information system. Based on the real-time, sudden, synchronous, and discrete characteristics of the control process of the welding flexible processing unit, references [23-24] used discrete event dynamics theory to analyze and study it. Combining the characteristics of WFMC sensing and control information flow, Petri net theory was introduced into the welding process. The unit information flow modeling and control method were systematically discussed. At the same time, the theoretical characteristics of the Petri net model were analyzed. On this basis, the implementation of optimized scheduling of the central monitoring system of the welding flexible processing unit was studied. 7. Conclusion The flexible manufacturing system for welding includes various robots (welding robots, handling robots, etc.), various tooling equipment and various production resources. They are typical representatives of multi-agent systems (MAS) in distributed artificial intelligence (DAI). In the flexible manufacturing system for welding, local reasoning and decision-making at different functional nodes are very important. Therefore, it is meaningful to introduce the concept of Agent at these nodes and use the mutual coordination between multiple Agents to achieve coordination between actual production equipment or production tasks. In the flexible manufacturing system for welding, robots, positioners, sensors, coordinators and other units are regarded as multi-agents working in coordination in the system. It is meaningful to study the technology of interconnecting, sharing resources, cooperating and completing one or more control tasks between intelligent subsystems that are dispersed, loosely coupled, geographically or logically dispersed and relatively independent in the network environment, so as to realize intelligent coordinated control of the flexible manufacturing system for welding [26]. References 1. Trailer. Manufacturer Depends on Robotic Welding to Boast Production. Welding Journal. 1995, 74(7): 49-51 2. West Carrillton. Robot Assure Quality for Auto Seat Manufacturer. Welding Journal. 1995, 74(8): 57-59 3. JDLane. Roboric Welding State of the Art. Robotic Welding—International Trends in Manufacturing Technology. IFS (Publications) Ltd. UK. 1987: 1-10 4. Gerhard Teubel. Experience+Application Up-Date: Automation of AW-Operations Using Robot-Technology. Welding in the World. 1994, 34: 75-84 5. Paul Drews., Gunther Starke. Welding in The Century of Information Technology. Welding in the World. 1994, 34: 1-20 6. Bob Irving. Sensor and Control Continue to Close the Loop in Arc Welding. Welding Journal. 1999, 78(5): 31-36 7. Pan Jiluan. Prospect of Welding Science Research in the 21st Century. Proceedings of the 9th National Welding Conference. 1999, (1): D-001-D-017 8. Song Tianhu, Li Minxian. The Development of Advanced Manufacturing Technology and the Future of Welding Technology. Proceedings of the 8th National Welding Conference. 1997 9. Lin Shangyang, Chen Shanben, et al. Welding Robots and Their Applications. Machinery Industry Press, August 2000. 10. Jiang Xinsong. Discussion on the Future Development Direction of Robot Technology. Robot. 1996, 18(5): 287-291 11. Wu Lin, Chen Shanben, et al. Intelligent Welding Technology. National Defense Industry Press, July 2000. 12. Chen Shanben, Wu Lin. An overview of research and application of robotic welding technology in my country. Welding. 1997, 11/12: 2-5 13. Chen Shanben, Wu Lin, et al., Research on Intelligent Control Technology and System for Spatial Weld Quality in Robot Welding, Technical Summary Report of Key Project of National Natural Science Foundation of China, 2001. 6. 14. Y.Li and JEMiddle. Machine Vision Analysis of Welding Region and Its Application to Seam Tracking in Arc Welding. Journal of Engineering Manufacture. 1993, No. 5: 275-283 15. L.Wu, K. Cui, SB Chen, 2000, "Redundancy Coordination of Multiple Robotic Devices for Welding through Genetic Algorithm," ROBOTICA, vol.18, No.6, pp669-676. (SCI, EI). 16. JS Tian, ​​Lin WU, et al., A General Algorithm of Rotating/tilting Positioner Inverse Kinematics for Robotic Arc Welding Off-line Programming , J.of Materials Science & Technology, 2001, 17(1): 161-162 17. R.Kovacevic, YMZhangand L.Li. Monitoring of Weld Joint Penetration Based on Weld Pool Geometrical Appearance. Welding Journal. 1996, Vol. 75(10): 317s-329s 18. RWRichardson and DAGutow. Coaxial Arc Weld Pool Viewing for Process Monitoring and Control. Welding Journal. 1984, Vol. 63(3): 43-50 19. SB Chen, YJLou, L.Wu, DB Zhao, 2000, "Intelligent Methodology for Measuring, Modeling, Control of Dynamic Process During Pulsed GTAW ——-PrtI Based-on-Welding,” Welding J . vol.79, No. 6, pp151s-163s. 20、SB Chen, DB Zhao, L. Wu, YJLou, 2000, “Intelligent Methodology for Measuring, Modeling, Control of Dynamic Process During Pulsed GTAW - PrtII Butt Welding,” Welding J. vol.79, No.6, pp164s-174s. 21、DBZhao, SBChen, L.Wu, et al., 2001, “Intelligent Control for the Double-sided Shape of the Weld Pool in Pulsed GTAW with Wire Filler,” Welding J. vol.80, No.11: 253s-260s. 22、Chen Shanben, Chen Wenjie, Lin Tao, Identification Model of Dynamic Process of Pulsed GTAW Molten Pool, Journal of Welding, vol.22, No.3, 5-8, 2001. 23. T. Qiu, SB Chen, YT Wang, L. Wu, 2001, “Modeling and Control of Welding Flexible Manufacturing Cell Using Petri Net,” J. of Material Science and Technology, vol.17, No.1, pp 185-186 (SCI). 24. T. Qiu, SB Chen, YT Wang, L. Wu. Information Flow Analysis and Petri Net-based Modeling for Welding Flexible Manufacturing Cell. SPIE International Symposium on Intelligent Systems and Advanced Manufacturing, Boston, Massachusetts USA, 2000, 11 (EI, ISTP). 25. Zheng Dazhong, Zheng Yingping. Theory of Discrete Event Dynamic Systems: Current Status and Prospect. Acta Automatica Sinica. 1992, 18(2): 129-142. 26. Wang Yuechao, Tan Dalong. Research Status and Development of Collaborative Robotics. Robot. 1998, 20(1): 69-75 ＊Chen Shanben, Director of the Robotics and Automation Professional Committee of the Welding Society, Professor of Shanghai Jiaotong University