航空发动机性能图像纹理片段划分方法  

Texture Segmentation Method For Aeroengine Performance Image

作　　者：蔡舒妤[1] 史涛 师利中[1] CAI Shu-yu;SHI Tao;SHI Li-zhong(College of Aeronautical Engineering,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学航空工程学院,天津300300

出　　处：《航空发动机》2023年第4期86-94,共9页Aeroengine

摘　　要：为了解决深度学习在航空发动机性能异常检测中出现的数据不平衡的问题,在高频次纹理片段对应性能稳定状态、低频次纹理片段对应异常波动状态的合理假设下,提出了一种航空发动机性能图像纹理片段划分方法。通过研究性能数据空间与RGB彩色空间的映射,提出高维性能数据图像化。引入分形盒维理论,定义纹理特征维数表征性能图像纹理特征,提出基于高频次生长树的图像纹理片段划分方法,实现性能图像按纹理出现频次的精细划分,建立正常样本和异常样本数量接近的数据集。选取多个航段发动机性能数据,与时频图和采样法生成的数据集进行对比验证。验证结果表明:该方法可以生成数量相当且真实客观的正常样本和异常样本,在ResNet50模型上的准确率达91.72%,减小了数据不平衡对准确率的影响。In order to solve the problem of data imbalance in aeroengine performance anomaly detection by deep learning,a texture segmentation method for aeroengine performance images was proposed under the reasonable assumption that high-frequency texture segments correspond to stable performance states and low-frequency texture segments correspond to abnormal fluctuation states.By studying the mapping between performance data space and RGB color space,the high dimensional performance data pictorialization was proposed.The fractal box dimension theory was introduced to define the texture feature dimension to characterize the performance image texture features,and the image texture segmentation method based on a high-frequency growth tree was proposed to realize the fine segmentation of the performance image by the frequency of texture occurrence,achieving the balance of normal and abnormal samples in the data set.The engine performance data of several flight segments were selected and compared with the data sets generated by the time-frequency plots and sampling methods for verification.The verification results show that the method can generate a considerable number of realistic and objective normal samples and abnormal samples,with an accuracy of 91.72%on the ResNet50 model,reducing the impact of data imbalance on accuracy.

关 键 词：性能图像 纹理特征维数 高频次生长树 图像纹理片段划分 深度学习 航空发动机 

分 类 号：V239[航空宇航科学与技术—航空宇航推进理论与工程]