Research On Quantitative Recognition For Bonding Flaw Of Composite Marterial Based On Feature Weithting SVM

Bonding technology is widely used in modern industry, especially in defense industry. But various catastrophic accidents will happen if there are de-bonding, bonding strength decreased or the strength does not meet the design requirements. Therefore, the safety and quality of bonding structure have been the issue of widely concern and an effective method for detection and recognition is needed. The safty and quality of bonding structure besomes a researching field and issue which is pervasively concerned by researchers coming form all over the word.At present, many researches of de-bonding detection are focused on qualitative analysis, but research on quantitative recognition is still in primitive stage. This paper devotes research on quantitative recognition for de-bonding of steel-rubber materials. The traditional method which resolves this problem is fuzzy neural network. However, neural network method has its inherent disadvantages:the structure of network is not stable and local minimum points are easy to be fallen. Support Vector Machine (SVM) is a new statistical technology to aim at tiny sample; it calculates globally optimal solution by means of searching the structure whose risk is minimal. This algorithm has been researched pervasively since it is proposed, but it is still not applied to de-bonding detection. Quantitative recognition for de-bonding of composite plate is a problem about tiny sample. This paper proposes SVM to recognize de-bonding degree of composite plate quantitatively in 10-level on the basis of the effective features.Traditional SVM is proposed to resolve 2-classification problem, classification of quantitative recognition for 10 de-bonding degree continuous is a typical multi-classification problem, and so multi-classification SVM model is established in this paper. Meantime, K-CV parameter searching method is applied to determining parameters of multi-classification SVM model. In addition, the three features extracted are nonlinear, so the result of recognition is not precise enough if using traditional linear SVM. In order to improve effect of recognition, this paper proposes nonlinear feature weighing SVM method to resolve non-linear problem, analyzes six methods of feature measurement and chooses the weigh factors calculated by Relief-F method for each features vector. From the comparisive results, it can be seen that feature weighing SVM improves the precision of quantitative recognition; this method provides a more precise criterion for quantitative recognition of de-bonding in composite plate, and strengthens the foundation of automated detection for de-bonding in composite plate.