Research And Implementation Of Aviation Experimental Data Integration And Analysis Platform

Building a standardized platform for collecting,storing and integrating aviation experimental data and integrating the latest AI technology is a new way for aviation industry practitioners to conduct experimental analysis.In recent years,more and more experiments have been carried out on aero-engines.Due to the high cost of aero-enginerelated experiments,the relevant experimental data show the characteristics of small samples.Accurate measurement of the total temperature of engine intake air is the focus of many experiments.The main work of this thesis includes:(1)Through the analysis of aviation experiment data,the aviation experiment data integration and analysis platform is designed and implemented.The data service layer in the platform is equivalent to the data warehouse module,which devises a collection,storage,and computation scheme based on the data warehouse for aviation experiment data from multiple sources and varying types;The platform layer of the platform corresponds to the client module,which provides users with data acquisition and storage function,data calculation function,data set generation function,algorithm execution function,platform monitoring and alarm function,data visualization function,etc.The business application layer of the platform corresponds to the algorithm execution module,which realizes the whole process of algorithm deployment,training and prediction,and supports the intelligent analysis function of the aviation experiment data integration and analysis platform.(2)In view of the small sample scenarios often encountered in aviation experiments,this thesis proposes a MAML network based on memory dictionary and parameter hierarchy,which is called ML-MAML for short;The traditional MAML network parameters are divided into fixed layer parameters and personality layer parameters.The fixed layer parameters are frozen during meta training,and only the personality layer parameters are fine-tuned.At the same time,in order to make the initialization parameters during fine-tuning conform to the current task characteristics,the memory dictionary module is used to store and match the appropriate personality layer parameters.Compared with the small sample learning method of MAML,Acc has increased by 8% in the setting of 5-way 5-shot on the mini Image Net dataset.(3)ML-MAML is used to solve the calibration application of total inlet temperature of aero-engine in the small sample scenario.The experimental results demonstrate that ML-MAML achieves an accuracy of 80.58%,surpassing the original MAML method by 6.76%.