Abstract: The quantitative method of magnetic anomaly of superalloy surface defects based on the parameter optimization of SVM was of great significance to solve the engineering problems of surface crack detection for turbine disks. A sample database was constructed with 16 preset magnetic anomaly characteristic values of defects obtained by weak magnetic detection, a SVM prediction model was established to invert and quantify of defect sizes. The validity of parameter optimization and the SVM model were verified by the specimens with known defects. The results show that after parameter optimization, the prediction results of length, width and depth are improved compared with the prediction results of default parameters, especially the inversion effectiveness of length and depth are significantly improved, and the prediction accuracy of genetic algorithm is better than that of cross-validation method. When the magnetic permeability of the defect and the base material is very different, such as ferromagnetic inclusions on the surfaces of the superalloy, the characteristic value amplitude of the magnetic anomaly is more larger. After the characteristic value amplitude and area are halved, the accuracy of the results is improved by more than 20%. The SVM model with genetic algorithm parameter optimization still shows good predictive ability for data outside the sample database, and the prediction accuracy is close to 85%.

Key words: superalloy, weak magnetic detection, defect inversion, parameter optimization, support vector machine(SVM)

摘要： 研究基于支持向量机参数优化的高温合金表面缺陷磁异常定量方法，对解决涡轮盘表面裂纹检测的工程问题具有重要意义。以弱磁检测得到的16个预置表面缺陷磁异常特征值构成样本库，建立支持向量机预测模型对缺陷尺寸进行反演定量，通过已知缺陷试件验证参数优化和支持向量机模型的有效性。研究结果表明，参数优化后长度、宽度、深度的预测结果比默认参数的预测结果都有提高，尤其是长度和深度的反演效果有显著提高，且遗传算法比交叉验证法的预测精度更高；当缺陷与母材的磁导率差异较大时（如高温合金表面的铁磁性夹杂），磁异常特征值幅值偏大，对特征值幅值和面积减半后进行反演，得到的结果准确度提高了20%以上；遗传算法参数优化的支持向量机模型对样本库之外的数据仍表现出较好的预测能力，预测准确度接近85%。

关键词: 高温合金, 弱磁检测, 缺陷反演, 参数优化, 支持向量机