基于概率支持向量机原理的超声缺陷识别模型研究

为了提高大型零件超声波探伤过程中的缺陷辨识能力，提出一种基于概率支持向量机原理，结合经验模式分解和DS证据理论，采用多探头检测的一种超声缺陷识别模型。首先，对每个探头检测的含有缺陷的信号运用经验模式分解法提取信号特征；其次，利用支持向量机来进行缺陷识别，并采用最大后验概率策略来处理传统支持向量机的输出，得到每个探头检测到的缺陷的概率支持度；最后，采用 DS证据理论得出最终的缺陷类型。结果表明，该模型克服了传统的支持向量机在处理多类问题时其硬判决输出限制后续数据处理的缺陷，同时避免了主观判断，提高了识别精度和准确率。与神经网络结合DS证据理论模型和单探头多级二类支持向量机模型进行了对比分析，论证了本模型的优越性。

Recognition model of Ultrasonic Flaws Based on PSVM

In order to improve the ability of flaws identification in ultrasonic testing, a flaw-recognition model based on Probabilistic Support Vector Machine, combined with Empirical Mode Decomposition and Dempster-Shafer Evidence Theory, was proposed to test on a large rotor with multi-ultrasonic sensors. Firstly, the characters of test signal were extracted with theory of Empirical Mode Decomposition. Secondly, a step forward was added to the output of the SVM classifiers to choose the category with a maximal posteriori probability, thus, an algorithm model of Probabilistic Support Vector Machine was presented. The outputs of Probabilistic Support Vector Machine were just the support degree of ultrasonic flaws. Lastly, the results of ultrasonic defects recognition were obtained with Dempster-Shafer evidence theory. Results showed that the proposed model in this paper overcame the limitation that the outputs of the traditional support vector machines were un-calibrated and should not be used to determine the category when a multi-class problem was presented. Comparison demonstrated that this model had a better performance in improving the recognition accuracy and nicety ratio of defects identification than the model of NN combined with DS and the model of SVM with odd sensor.