Mar.2005, Volume 2, No.3 (Serial No.4)
Journal of Communication and Computer, ISSN1548-7709,USA
Web-Based Learning and Fault Diagnostic System PAN Feng, ZHU Jianghua School of Information & Control Engineering, Southern Yangtze University, Wuxi, Jiangsu 214036,China Abstract: This paper presents a novel knowledge-based multi-agent system for remote fault diagnosis, which is composed of learning and diagnostic agents (LDAs), machine agents (MAs) and a central management agent (CMA). Machines are remotely diagnosed by the LDAs through the communication channels between the MAs and the LDAs. When faults that cannot be solved by the present knowledge base occur, the LDAS can acquire new knowledge, translate it into rules using a rule builder, and update the rules into the CKB(Central Knowledge Base). The CKB will become mature through a continuous learning process. A prototype system has been developed and used for remote fault diagnostics of tool wear in computer numerically controlled (CNC) machining. Key words: Expert Systems; Fault Diagnosis; Knowledge Acquisition; Multi-agent Systems *
1. Introduction Manufacturing industries are facing serious structural problems brought about by rapid developments of overseas activities and manufacturing factories. Factories located in different regions must be coordinated through the use of state-of-the-art information technologies to ensure consistent product quality. As a result, manufacturing activities can be integrated and monitored in many regions and countries. The performance of a machine can be monitored and accessed from anywhere in the world. In addition, information on productivity, diagnostics, and training of manufacturing systems could be shared among partners at different locations. The development of a remote diagnostics system will provide manufacturers and users with greater flexibility for conducting manufacturing activities. At present, remote diagnosis through the Internet PAN Feng, professor. He graduated from Southern Yangtze University in 1984. His research interests are distributed processing, grid computing, network security and e-commence. Tel: +(86)-510-8883618. E-mail:
[email protected]. 44
does not have many practical applications in shop-floor manufacturing. In addition, many of the knowledge bases used in the diagnostic systems are traditional rule bases. That is, before a knowledge base is put into use, it is built largely enough to include as much knowledge as possible. During the working process, new knowledge that may be added to one knowledge base will still be unknown to the others. If there was a central learning knowledge base that could be shared by all the users at the different factories, it is possible to collect as much knowledge as possible through the networks, acquire knowledge once and use it in all the factories. The repetition of the knowledge acquisition process will be reduced. With the use of this system, the knowledge base will grow and mature. Based on this consideration, a web-based learning knowledge- based system is developed in this project. Multi-agent technology is used in the system. Learning and diagnostic agents (LDAs) are created to support multi-user remote diagnosis (working machines located in different places), and learn new knowledge through the faults that occur at all the sites. The central management agent (CMA) is in charge of updating the knowledge in the central knowledge base (CKB). The machine agent (MA) keeps records of the machine working status. These agents cooperate to realize remote monitoring, fault diagnosis, and on-line knowledge acquisition.
2. Architecture of the Multi-Agent Remote Diagnostic System The framework of a remote fault diagnostic system with learning capability is shown in Fig. 1. According to the differences in location and functionality, this system
Web-Based Learning and Fault Diagnostic System
is divided into three modules: the central management system (CMS), learning and diagnostic agents (LDAs) and the remote machine site (RMS). Each module dedicates to some specific functions. Using agent technology in this system, these three functional modules are wrapped into three corresponding agents –the CMA, the LDA and the MA. CMS is central management system; CMA is central management agent; LDS is learning and diagnostic system; LDA is learning and diagnostic agent; RMS is remote machine site; MA is machine agent. 2.1 Machine Site In this system, the machines to be diagnosed are located at different locations from the CMA. The machines are wrapped with the agent software into MAs. These MAs are in charge of sending requests to the CMA, sending real-time status signals to the LDA, and receiving messages from the other agents. Theoretically, supposing the network transmission speed is fast enough, and the hardware and software can satisfy the response time requirements of the diagnostic system, the machines to be diagnosed on-line or off-line
can be located around the world. For on-line fault diagnosis, it is essential to locate the machines inside an Intranet to ensure a quicker response. In this project, it is assumed that the network transmission speed is fast enough to allow globally distributed machines to be diagnosed remotely in real time. In this system, it is assumed that the overall number of machines at the remote sites is finite, and the types of machines are known. However, at any specific time, the number of working machines is not definite, and the corresponding LDAs will be initiated to implement the fault diagnosis and learning for these machines. 2.2 Diagnostic and Learning Agent The module with the functions of fault diagnosis and learning is encapsulated into a LDA. In the prototype, one LDA is created to serve one machine at a specific machine site. Through communication between the remote machine and the LDA, real-time signals are transmitted to the LDA. The LDA monitors these machines and diagnoses the faults in real time. At the same time, new knowledge is acquired using a learning algorithm through the fault diagnosis process.
CMS/CMA Valuable rules
Central Knowledge Base
Update Records
update request
LDS/LDA
new rules
Learning Engine
rules Inference Engine
Working Memory
report about new rules
rules
Knowledge Base data
Example Base
Report Base
Rule Builder
update request user input data
user input data RMS/MA Machine With Multi-sensor
Status Records Database
Fig. 1 System Architecture Notes: CMS, central management system; CMA, central management agent; LD S, learning and diagnostic system; LDA, learning and diagnostic agent; RMS, remote machine site; MA, machine agent.
45
Web-Based Learning and Fault Diagnostic System
(5) finish User’ s Machine Agent
Central Management System
(1) request (3) ready (6) finish (death)
(4) connect
(2)Initiate (birth)
User’ s Machine Agent Fig. 2 Working Process of this S ystem
2.3 Central Management Agent The CMA consists of a central knowledge base (CKB), an example base (EB) and a report base (RB). The CMA has the responsibility for updating new rules into the CKB, and maintaining it. After the learning processes have been implemented by the LDA, the newly created knowledge and the corresponding reports (explanations), from the learning engine in the LDA, are sent to the CMA, which stores them in its EB and RB. When the CMA receives a message from the LDA that there is no solution in the current LDA’s knowledge base for a fault in a machine, the CMA will load new knowledge from the EB and send it to the LDA. Each time new knowledge is used in diagnosing a fault on the machine, the confidence factor for this new knowledge, which is a factor to test the typicality of that knowledge, will be increased by one count. This knowledge is added to the CKB in the CMA when its confidence factor is high enough. The knowledge base of the LDA is updated periodically with the new knowledge in the CKB. The system engineer sets an upper limit for the confidence factor. With the operation of this system, the CKB will grow and become very large. Hence, the long-term maintenance becomes an important issue for this system.
3. System Operation 3.1 Working Process Stationary agents are used in the prototype. The 46
LDAs are located at different sites from the machines. The entire working process is shown in Fig. 2. When any one of the remote machines in this network begins to work, the MA sends out a “request” message to the CMA. After receiving a request from the MA, the CMA initiates a LDA for that machine, and sends the specific name of the LDA and a “ready” message to that MA. By calling the name of the LDA, the MA at the remote place establishes a connection with the corresponding LDA. The signals of the working status from the remote machine can be transmitted to the LDA in real time. The LDA monitors and tests the working condition of the machine on-line, to make sure that no abnormal condition appears, and the knowledge learned during the learning process is also transferred through the communication channels. When the machine finishes the jobs and is to be shut down, the MA sends out a “finish” message to the CMA, and the LDA changes into a “waiting” mode and waits for a new arrangement (new connection to some machines) from the CMA. Although the number of working machines at a specific time is not finite, the number of average working machines at a certain time can be estimated. Thus, the number of LDAs in the system can be set equal to the average number of working machines. If there are not enough LDAs, the CMA can initiate a new LDA easily (an inference engine, a learning engine and a rule builder are created). 3.2 Diagnostic Process
Web-Based Learning and Fault Diagnostic System
In this stationary multi-agent system, when a LDA is initiated, its knowledge base loads rules from the CKB in the CMA. Only the rules that are correlated to that machine are loaded. The diagnostic process in a LDA is shown in Fig. 3. Wait for data Signal input Y Normal?
Wait for next data N
Read working status data from working memory and reason in inference Engine
Y Implement the solutions
Find solution to faults?
CMA, in order to load the related rules into the rule base in the LDA. The inference engine will reason again with the expanded knowledge base. An answer may be found using the added rules in the LDA. The corresponding confidence factor of the fired rule is increased by one count at the same time. After finishing these processes, the added rules will be removed from the knowledge base in the LDA and sent back to the EB. 3.3 Learning Process The learning process of the LDA begins when a solution cannot be found using the rules in the EB. The fault is new to the diagnostic system, and the knowledge acquisition process must begin, which needs help from the domain experts in this prototype system. The learning process is shown in Fig. 4.
N
Abnormal status
Load rules from Example Base
Domain expert Y
Find solution to faults?
Increase the confidence factor
N Domain Expert Fig. 3
Find the problem directly Rule Builder
Collect data from training NN Extract rule by NN
The Diagnostic Process
The LDA receives the status data from the machines in real time. The data is sent to its working memory. The inference engine loads the data from the working memory, and makes use of the rules from the knowledge base for reasoning. When the input signals are abnormal, the inference engine sends out an alarm message or some control signals to control the machine remotely and avoid the danger. If the fault is not new to the knowledge base, this system will take action immediately to remove it. However, it is possible that the faults are unknown to the diagnostic system, i.e. the rule base does not have sufficient knowledge to correct the faults. In such a condition, the LDA will request help from the CMA. The CMA will connect the LDA’s knowledge base with the corresponding parts of the EB in the
Create new rule using the Rule Builder
Refine the learnt rules in the Learning Engine Update Example Base periodically Update Central Knowledge Base periodically Update Knowledge Base in every LDA Fig. 4 The Learning Procedure
The domain experts have two responsibilities: first, when the faults cannot be removed automatically by the system, the domain experts must go to the shop floor, check the records of the working status in the status records database, find the faults and propose
47
Web-Based Learning and Fault Diagnostic System
reasonable solutions. After that, the domain experts assume the second responsibility: learning. They input the new knowledge in the form of a “rule” into the rule builder. The rule builder translates this original rule into the expert system language. At the same time, the domain experts should submit a short report to explain the new rule and give a detailed introduction to the fault and the solution. These documents are very useful for the maintenance of the CKB. The newly created rules from the rule builder are of the same form as the rules in the CKB. However, these new rules may be partly repeated, or conflict with the rules in the CKB. With the rules provided by the rule builder, newly refined rules are created in the learning engine based on the rules in the knowledge base of the LDA, and thus the learning process in the LDA is completed.
4. System Implementation Java is used to develop the system, as it is a robust and reliable programming language that is platform-independent and Internet-supported. It is the best choice for developing a multi-agent system running on the Internet. Sensors are necessary to monitor the working status of the machines, such as force, acoustic emission, power consumption, etc. The data to be transmitted for diagnosis is compressed before being sent out, in order to decrease the transmission loads. In practice, the diagnostic process may need more information about the machine status besides the signal inputs from the sensors. Multimedia technology may be used to acquire more information. For example, video-conferencing packages [1] can provide dynamic image transferring from the remote machines. In addition, in certain conditions, the operators at the machine sites can be asked to provide the necessary information to the LDA directly through the agent telecommunication channel. The transmission of the machine working status
48
data is implemented in a TCP/IP client/server structure using Java. The machine works as a signal server, and a server socket is created using Java to send out signals from the sensors in real time. On the LDA side, a socket is coded into the agent to receive the signals from the machine. The learning function is an important feature of this system. As discussed before, there are many different algorithms to realize the learning of new rules, such as extracting rules from trained artificial neural networks [2], learning rules from experience using genetic algorithms [3], and extracting rules from databases using rough set theory [4]. In this project, a learning method based on the theory reported by Guan and Braham [5] is used. This method identifies all the possible paths of fault propagation for any possible failure sources through the representation of the physical connectivity of the devices under diagnosis. The structure is classified into many sub-devices that are interconnected. The interconnections between the sub-devices provide information on the possible paths of fault propagation, so that it is possible to trace the faults back to the responsible root causes of the faults. The CKB is classified into three parts to include some of the common machining operations in this system. It includes the turning rule base, milling rule base, and the drilling rule base. Each rule base has its own rules for the corresponding machining process. When a machining process, such as the turning process, is initiated on a remote site, the CMA initiates a LDA (an inference engine, a learning engine and a rule builder are created). The knowledge base in the LDA loads all the rules from the turning rule base. Each rule has a rule name, which is composed of a class name and an identification name. The class name is the rule base name. For example, rules in the turning rule base have the class name TURNING. The identification name is the name that stands for the rule content. When a new rule is extracted through the learning process, its rule name is extracted from the
Web-Based Learning and Fault Diagnostic System
report files. Every report file generated from the learning process, has the same class name and the same identification name as the name of the corresponding rule. The EB and RB are also classified into three parts: turning, milling and drilling.
5. Case Studies
the sys tem. The CMA is located on PC 1, on which the router is running. Here, the LDA is installed on PC 1. It can also be located on a different PC from the CMA. PC2 is used as the MA. These two PCs are connected through the Internet. The connection between the three parts is shown in Fig. 5.
5.1 Monitoring and Diagnosis of Tool Wear In machining operations, the cutting tool usually performs under severe conditions of high temperatures (800–1000 ℃ ) and high forces (2500–3000N).
Central Management Agent (PC1)
Gradually, the tool will lose its capability to produce the intended objective of cutting. Failures to detect tool breakdown may result in damage to the work piece and the machine and poor work piece quality [6–8]. Li, etc [9] recently reported a neural network model with fuzzy logic as a hybrid-learning model for tool wear monitoring in drilling operations. The tool life can be estimated in different ways. Recent studies on the dynamic cutting force have shown that on-line tool wear monitoring may be feasible. The onset of tool failure can be predicted by determining the threshold value of the percentage drop in the dynamic force, from its maximum amplitude, before the tool fails. This is a good indicator for predicting tool failure and it can be incorporated into the software program for on-line tool wear monitoring. In addition, the value of the dynamic force does not always increase monotonically with time or tool wear. These fluctuations of the dynamic force signals can be misinterpreted as a change in the trend when the effect may only be transient. The change in the trend of the dynamic force can be better assessed by setting a second condition that checks the gradient of the dynamic force curve. These two suggested conditions must be met to indicate the onset of tool failure. 5.2 Case Study Currently, a simple prototype of the web-based knowledge-based system for tool wear fault diagnosis has been constructed. Two PCs are used to implement
1
Learning and Diagnostic Agent (PC1)
5
2
3 4
Machine Agent (PC2) 6 Machine Fig. 5 Structure of the Prototype S ystem
Sensors could also be used to acquire other information, such as temperature, and acoustic emission. The signals from the lathe are processed on the MA, and sent out through a server socket coded in the MA. The whole system is initiated by the following steps: Step 1. Run the Agent Message Router. Step 2. Initiate Central Management Agent. Step 3. Run the Machine Agent. Step 4. Ready for the Preparation. Step 5. Create Connection between MA and DLA. Step 6. Remotely Diagnose the Machine.
6. Conclusions Case study demonstrates the capability of real-time on-line tool wear monitoring in the formulated web-based and agent-based architecture. Signals from machines located remotely from the server are transmitted via the Internet to the server where the CKB is located for diagnosis, and solutions to these faults are transmitted in real time to the remote machines to correct the faults. In the prototype
49
Web-Based Learning and Fault Diagnostic System
developed in this work, the stationary agent technology is used. It is possible to apply the mobile agent technology in this structure. However, a compromise has to be made between the network safety and the traffic. In addition, special requirements such as the central control of the subsystems should be taken into consideration in the design. This system has several advantages. Through the central management system, the working conditions of all the distributed shop floors can be monitored and diagnosed in real time. Located at the site of the central management system, the system manager can acquire the information of all the remote machines. Secondly, in this architecture, the shop floors share the same CKB. This saves the cost of maintaining the knowledge base at each site. The updating of the CKB is easier and less costly. All the machines are connected through the Internet, which facilitates the knowledge acquisition process. Each useful rule acquired from one site can be used by the other sites, which have similar machines. This takes advantage of the web-based technology in knowledge acquisition. Finally, in this system, some methods are introduced to improve the maintainability of the CKB, such as modularity, graphical viewer, etc. References: [1] P.Wayner, New Videophones Starved for Bandwidth, Byte, 21(5), pp. 125–128, 1996. [2] R. Andrews, J. Diederich and A. B. Tickle, Survey and
50
[3]
[4]
[5]
[6]
[7]
[8]
[9]
Critique of Techniques for Extracting Rules from Trained Artificial Neural Networks, Knowledge-Based Systems, 8(6), pp. 373–389, 1995. F. J. Garrido and M. A. Sanz-Bobi, Learning Rules from the Experience of an Expert System Using Genetic Algorithms, Second International Conference on Genetic Algorithms in Engineering Systems: Innovations and Applications, conference publication no. 446, pp. 226–231, 1997. S. Tsumoto, Discovery of Rules for Medical Expert System-rough Set Approach, Proceedings of Third International Conference on Computational Intelligence and Multimedia Applications, pp. 212–216, 1999. J. Guan and J. H. Braham, An Integrated Approach for Fault Diagnosis with Learning, Computers in Industry, 32, pp. 33–51, 1996. T. Nakajima, J. Ahn and T. Sata, Tool Breakage Monitoring by Means of Fluctuations in Spindle Rotational Speed, Annals CIRP, 36(1), pp. 49–52, 1987. S. B. Rao, Metal Cutting Machine Tool Design –a review, Transactions ASME, Journal of Manufacturing Science and Engineering, 119(4), pp. 713–716, 1997. K. Danai and A. G. Ulsoy, Dynamic State Model for On-line Tool wear Estimation in Turning, Transactions ASME, Journal of Engineering for Industry, 109(4), pp. 396–399, 1987. X. Li, S. Dong and P. K. Venuvinod, Hybrid Learning for Tool Wear Monitoring, International Journal of Advanced Manufacturing Technology, 16, pp. 303–307, 2000. (Editors: WANG Nai, XIE Bei, ZENG Dan)
