Research On Rough Set Knowledge Modeling For Expert System Prediction Of Weld Seam Quality

Welding automation is a trend of welding technical development. Be-cause of the high complexity of arc welding process, it is hard to obtain accurate mathematical models, which limits the further application of tradi-tional control method. Recently, more attention has been paid to intelligent control for its adaptability to complex processes. Intelligent modeling is usu-ally essential and significant to the design of intelligent control system. Re-searches are focused on fuzzy set method, neural network method or their hybrid method for welding process's intelligent modeling. However, these methods have some disadvantages that are hard to overcome by themselves. Rough set (RS) based modeling method, being a relative new modeling method, has been used in welding field and has shown its adaptability. Ac-cording to request and character of expert system prediction for weld seam quality, a knowledge modeling method based on rough set was proposed in this dissertation.This modeling method included four steps: raw data acquisition, data preprocessing, RS model reduction and knowledge reasoning. Discretization algorithms of data and uncertain reasoning strategy were studied in this dis-sertation deeply.(1) Research on discretization algorithms of dataIt was necessary when there were continuous data in decision table. In this dissertation, the evaluation standard for discretization algorithms in welding were decided at first. Then three discretization methods are analyzed, and they are equal-width intervals discretization, discretization based on Ko-honen net (it can also been called self-organizing mapping) and intervals combination discretization based on the importance of attributes. At the same time, Kohonen net is improved to make it more suitable for discretization in welding. On the basis of summarizing these two kinds of discretization methods and getting rid of their disadvantages, a new method called secondly discretization is brought forward, which means that on the base of equal-width intervals discretization or the discretization based on modified Kohonen net, intervals combination discretization based on importance of at-tributes is used. Experiment showed that prediction precision was advanced and prediction error is reduced,based on secondly discretization algorithms of modified Kohonen net.(2) Research on uncertain reasoning strategy This used the obtained RS model to predict the unseen samples'output and proper inference strategy will greatly improve the RS model's predictive ability. Because of the complexity and uncertainty of welding, uncertain rea-soning was requisite.The limitation of the already used reasoning method is pointed out, which lies in that the condition attributes of each rule were dealt with indist-inguishably in the process of reasoning, ignoring the significance of each condition attributes in a single rule. In consideration of this, a new reasoning method based on attributes significance is presented in this dissertation. The experiments have shown that this new method can increase the utilization ef-ficiency of knowledge base as well as the accuracy of prediction model.At last, database of mild steel CO2 welding was used to validate the knowledge modeling method based on rough set, the result showed that RS modeling method satisfied the real demand for expert system prediction of weld seam quality. According to research result in this dissertation, expert systems of monitoring for lack of weld and prediction for weld seam quality were constructed, and were applied in practice. Validity and applicability of the knowledge modeling based on rough set was proved in farther.