【正文】 on their expertise and efficiency. Data mining for machine fault diagnosis The unstructured textual data of faultcondition and checkpoint information of the customer service database provides useful machine service information. A data mining technique based on the integration of neural work, casebased reasoning, and rulebased reasoning has been applied to the customer service database to support intelligent machine fault diagnosis. It shows the framework of the data mining process. It consists of two major processes: the offline knowledge extraction process and the online fault diagnosis process. The first extracts knowledge from the customer service database to form a knowledge base that contains the neural work models and a rulebase. The neural work models and the rule base work within the CBR cycle to support the second, which uses the four stages of CBR cycle (retrieve, reuse, revise, and retain) to diagnose customer reported problems. It accepts user39。atindexed case database。 ., more than one data mining techniques. For example, Darwin from Thinking Machine Corp. supports neural works, regression tree (CART), kmeans algorithm, and case based reasoning for classification, prediction, and clustering functions. There are also some tools that only aim at a special data mining function. This provides 175。 Machine fault diagnosis Introduction Customer service support is being an integral part of most multinational manufacturing panies that manufacture and market expensive machines and electronic equipment. Many panies have a customer service department that provides installation, inspection, and maintenance support for their worldwide customers. Although most of these have some engineers to handle daytoday maintenance and smallscale troubleshooting, expert advice is often required from the manufacturing panies for more plex maintenance and repair jobs. Prompt response to a request is needed to maintain customer satisfaction. Therefore, a hotline service centre (or help desk) is usually set up to answer frequently encountered problems from the customers. Fig. 1 shows the workout in a traditional hotline service centre. The service centre is responsible for receiving reports on faulty machines or enquiries from customers via telephone calls. When a problem is reported, a service engineer will suggest a series of checkpoints for customers using the hotline advisory system. Such suggestions are based on past experience. This has been extracted from a Customer Service Database, which contains previous service records that are identical or similar to the current problem. The customer can then try to solve the problem and subsequently confirm, with the service centre, if the problem is resolved. If the problem still persists, the centre will dispatch a service engineer to the customer39。 it would search the unstructured customer service records for machine fault diagnosis. The proposed technique has been implemented to support intelligent fault diagnosis over the World Wide Web. Author Keywords: Data mining。 Knowledge discovery in databases。s premise for an onsite repair. During such trips, the service engineer will take past records of the customer39。edibility。 this is inefficient, especially for large case database. Other CBR systems use hierarchical indexing such as CART , decision trees , and although this performs efficient retrieval, building a hierarchical index needs the knowledge of an expert during the caseauthoring phase. The neural work approach provides an efficient learning capability when provided detailed examples. Neural works may be either supervised or unsupervised, depending on the method of training. It performs retrieval based on nearest neighbor matching, since it stores the weight vectors as the codebook or exemplar vector for the input patterns. The matching is based on a petitive process that determines the output unit that is the best match for the input vector, similar to the nearest neighbor rule. However, the search space in a neural work is greatly reduced because of the generalizations of knowledge through training. In contrast, the CBR systems need to store all the cases in the case database in order to perform accurate retrieval. The CBR systems that store only relevant cases for an efficient retrieval lack the accuracy as well as the learning feature. Thus, neural works are very suitable for case indexing and retrieval. Other data mining techniques include rulebased reasoning, fuzzy logic, geic algorithms, decision trees, inductive learning systems, and statistical pattern classification systems. In addition, hybrid approaches, such as hybr