- 上一篇：注塑模具单一浇口改善英文文献和中文翻译
- 下一篇：空调系统英文文献和中文翻译
2. Research background
2.1. Case adaptation in CBR
The case-based inference engine of CBR system solves new problems by retrieving and adapting several previous problem-solving experiences. So far, the research of retrieval method in CBR is quite mature, and designers can find several similar cases according to different types of design requirements with the help of retrieval methods. However, the difficult issue in successfully applying CBR is that the retrieved solution cannot be used directly, leading to the cause of the bottleneck problem of CBR. Many CBR systems normally utilize the most similar case as the only candidate in the adaptation process, which ignore the impacts of other useful similar cases]. This is the reason that they have limitations on providing precise solution in PMD. The study of adaptation by referring to more than one case(K > 1) is a more complicated issue, and several statistical methods have been employed in the research of multiple case adaptation since 1990s, such as the closet analogy, the equal mean (EM), the median, weighted mean (WM), and multivariate regression analysis (MRA). Nowadays they are also regarded as the baseline modes, whose advantages are the domain independent and easily to be implemented. Among them, the family of WM is the most widely used statistical method, which calculates the weighted average of the solution values of K similar cases. The weight of the solution value represent its significant to the adaptation result, and the assignment of weight value is based on the similarity value of retrieved case, i.e., wk j ¼ sk=Psk, where wk j refers to the weighting factor of jth solution value of kth retrieved case in WM model, and sk is the similarity value of kth retrieved case. Although the abovementioned statistical approaches are effective on case adaptation, limitations on precision also exist due to the fact that the information of similar cases has not yet been fully utilized. Later, the intelligent machine learning methods have been introduced to obtain the adapted values in CBR, which explore the utilization of inductive learning to acquire the differences between cases and their corresponding solutions and apply the acquired knowledge to implement automatic case adaptation. Different intelligent methods including neural networks, genetic algorithm, and support vector machine (SVM)], decision tree, and simulated annealing have been employed in this area. In mechanical design domain, Butdee and Jung used back-propagation neural network (BPNN) and radial basis function network respectively to perform case adaptation for aluminium extrusion die design and shadow mask design. Sharifi et al. constructed an adaptation model based on SVM to design servo system. Saridakis and Dentsoras researched the adaptation of the oscillating conveyor case through genetic optimization, and the retrieved solutions are used as the initial population. In general, intelligent methods can produce more accurate adaptation results than statistical methods on the whole. The drawback of intelligent methods lies on the sufficient case data needed in the process of case adaptation, and they could have expensive computational cost when working on data set with large number of case attributes.
2.2. Motivation and originality of this research
Subject to manufacturing costs and manufacturing capacity, the application of CBR for complex mechanical design uses data set with no more than 100 different types in one product family, so only limited cases can be collected for intelligent model training. Compared with intelligent methods, the statistical method is a relative suitable option for CBR adaptation in PMD, which does not need abundant computational data and is good for small data sets with multi-parameters. Therefore, this paper uses the WM as the basic model. On the other hand, the performance of WM model strongly relies on the importance of similar cases for the new problem. Therefore, the weighting factors play an essential role in improving the effectiveness of WM model. By far, most of studies use similarity values to figure out the weight values, and limited studies have ever attempted to improve case adaptation technique by addressing the study of weight distribution strategy. As we known, there have been no discussions on the effects of the mutual relationships between problem and solution features for the adaptation results. To address this issue, it is valuable to utilize certain data mining technique to elicit the problem–solution (PS) relational information from retrieved cases, and study whether the performance of WM is able to be improved with the help of PS relationship information. In this paper, the grey relational analysis (GRA) is used to abstract this implicit knowledge. The motivation for us to consider the GRA is its capability of solving the complicated interrelationships among designated feature values and it can deal with problems involving poor, insufficient, and uncertain information. Our paper will dedicate to integrate PS relational information into basic WM model in order to facilitate efficient case adaptation in application of CBR for PMD, and the new adaptation method is called as GRA-WM. Moreover, the empirical data are also collected to compare the adaptation performance of GRA-WM with that of commonly adaptation methods.