基于超声背散射信号特征优选策略的封严涂层完整性多参数SVR协同预测

可磨耗封严涂层质量控制是航空发动机“卡脖子”问题。涂层完整性多参数超声反演具有非线性、不确定性和多因素强耦合等特点，传统方法难以解决。拟采取“小样本数据+人工智能”方式，开展基于超声背散射信号特征优选策略的支持向量回归（SVR）协同预测研究。背散射信号辨识困难、单一类型参数信息有限，且不同超声参数与目标参数之间相关性迥异，为满足低维度、避免过拟合、与目标参量相关度高等输入特征要求，制订“背散射信号辨识准则+多域融合信号处理方法+显著性分析选择依据”的超声信号特征优选策略，为构建高质量输入特征空间提供科学依据。封严涂层完整性多参数之间关联性强，针对采用多个独立的单输出回归模型处理多输出问题存在整体预测性差、计算成本高的问题，选取复合核函数建立考虑多输出参量内在关联性的支持向量回归协同预测模型。借助启发式算法和交互检验技术优选模型参数，提高模型预测性能。研究工作具有重要科学价值和应用前景。

The quality control of the abradable seal coating is a "bottleneck" issue in aircraft engines. The multi-parameter inversion of seal coating integrity through traditional ultrasonic methods has the limitations of nonlinearity, uncertainty, and multi-feature strong coupling. A "small sample data + artificial intelligence" strategy is proposed to implement multi-parameter collaborative prediction of seal coating integrity based on feature optimization scheme of ultrasonic backscatter signals using support vector regression (SVR). The difficult identification of backscatter signals, limited information of a single parameter, and the different correlation between ultrasonic features and target parameters cause the difficulty of realizing low dimension, without overfitting, and high correlation with target parameters of input features. An feature optimization strategy of ultrasonic backscatter signal is created using " Identification rule of backscatter signals + Fusion method of multi-domain signal process + Selection criteria of significance analysis"，which provides a scientific basis for the identification, processing, and feature selection of backscatter signals. There is a strong correlation between the multiple parameters of seal coating integrity. Aim at the problem of poor predicted results and high calculation cost using traditional multiple single-target SVR models to solve multi-target problems, a multi-target SVR model constructed with hybrid kernel functions is created for multi-parameter collaborative prediction of seal coating integrity according to the multi-target’s correlations. The model parameters are globally optimized by combining heuristic algorithms and cross-validation methods to enhance the prediction performance of the multi-target SVR model. This program has significant scientific value and application prospect.