基于BPNN的封严涂层孔隙分布均匀性超声表征

Ultrasonic quantitative characterization of pore distribution uniformity of seal coating based on BPNN

可磨耗封严涂层的多相非均质特性导致超声波在其中传播行为复杂,涂层组成相分布均匀性的超声表征在有效超声特征提取、特征与均匀性关系描述等方面存在困难。针对这一现状,提出了BP神经网络(BPNN)结合超声信号小波变换技术定量表征可磨耗封严涂层孔隙分布均匀性的方法。基于统计学思想与随机场理论建立铝硅聚苯酯(AlSi-PHB)封严涂层随机多相介质模型,通过调节自相关长度参数改变模型的孔隙分布状态。采用面积分数多尺度分析技术获得定量描述涂层孔隙分布均匀性的绝对斜率与均匀性长度,提取超声仿真信号的时域、频域及小波分解时频域衰减系数,探究其与绝对斜率及均匀性长度之间的相关性,并作为BPNN的输入对绝对斜率与均匀性长度进行预测。结果显示:在孔隙率5%时,随自相关长度从11μm增大至26μm,绝对斜率从0.86减小至0.76,均匀性长度从550μm增大至6 785μm,表明孔隙分布均匀性降低;时频域衰减系数与均匀性参数间的相关性最显著,绝对斜率与均匀性长度的BPNN预测结果决定系数分别为0.97和0.98,在3种衰减系数预测结果中最优。所提方法可为定量表征非均质涂层组织分布均匀性提供有效途径。

The propagation behavior of ultrasonic wave is complex in multi-phase heterogeneous abradable seal coating,so it is difficult to extract effective ultrasonic characteristics and describe the relationship between the characteristics and microstructure distribution uniformity.The BP Neural Network(BPNN) combined with wavelet transform is proposed to quantitatively characterize pore distribution uniformity of abradable seal coating through the ultrasonic testing technique.Based on the statistical method and the random medium theory,random multiphase medium models for AlSi-Polyester(AlSi-PHB) seal coating were established,whose pore distributions were adjusted by varying autocorrelation lengths.Two uniformity parameters,absolute slope,and uniformity length were extracted to quantitatively describe pore distribution uniformity using the multi-scale analysis of area fraction.The time domain,frequency domain,and time-frequency domain attenuation coefficients were extracted from simulated ultrasonic signals.Correlations between absolute slope,uniformity length and the attenuation coefficients were explored.The BPNN with attenuation coefficients as the inputs was applied for training and prediction of absolute slope and uniformity length.The results show that when porosity was 5%,as autocorrelation lengths increased from 11 μm to 26 μm,absolute slope decreased from 0.86 to 0.76 and uniformity length increased from 550 μm to 6 785 μm,which indicates that the pore distribution uniformity gets worse.The time-frequency domain attenuation coefficients and the uniformity parameters were the most correlative,and the determination coefficients of the predicted absolute slope and uniformity length through BPNN were 0.97 and 0.98,respectively,which are the most accurate among the prediction results of three kinds of attenuation coefficients.The proposed method provides an effective way for quantitative characterization of distribution uniformity of multi-phase heterogeneous coatings.