Data-driven Based Anomaly Detection And State Recognition For Fuselage Join

Large aircraft components(fuselage,central wing,outer wing and tail wing,etc.)join is the process of assembling and connecting large components into the whole aircraft.The quality and efficiency of components join determine the final quality,manufacturing cost and cycle of the aircraft to a great degree.The process of automatic join of large aircraft components is complex,involving the collaborative work between multiple systems and the transition between various joining states.Any problem of any system or link in the joining process may lead to abnormal join and affect the quality of joining assembly.Intelligent perception and real-time analysis of joining anomalies and states in the automatic joining process of large aircraft components can help to find problems in time and take corresponding measures to ensure highquality join.Intelligent perception and real-time analysis are also important parts of the transition from digital assembly to intelligent assembly.In this paper,the joining process between the front fuselage and the middle fuselage of a certain type of aircraft is taken as the research object,and the method of data-driven based anomaly detection and state recognition for fuselage join is studied.The main contents are as follows:(1)The principle,technological process and characteristics of automatic join of large aircraft components are studied,and the joining process is divided into various joining states,such as empty station,joining preparation,posture adjustment and join,connection.The displacement and load data of positioner group during the fuselage join are collected and preprocessed,and the sensor damage,data acquisition system failure and other unforeseen anomalies during the automatic join are pointed out.(2)Principal component analysis is adopted to reduce dimensionality of the sequence data of fuselage join.Long Short-Term Memory network is designed and trained to predict the principal components of fuselage join.Unsupervised anomaly detection models based on isolated forest,local outlier factor and statistics are trained respectively by the difference between actual and predicted principal components.Then the process of anomaly detection using the anomaly detection model to deal with the detected data is designed.Multiple actual cases of fuselage join are used to verify and compare different anomaly detection methods,the results show that the anomaly detection model based on Long Short-Term Memory network and isolated forest gets the best result,which can perform efficient training and detection with extremely high precision and low false negative rate.(3)According to the process of large components join,the data of fuselage join are manually labeled,then an accurate and efficient automatic labeling method of join state is proposed.Several sequential/non-sequential state recognition models for fuselage join are proposed,which are validated and compared by cross-validation.The results show that compared with deep learning and other statistical machine learning methods,the state recognition model based on gradient boosting tree can recognize each joining state more accurately and has higher training efficiency.(4)Based on the method proposed above,the anomaly detection and state recognition system of fuselage join is designed and developed.Which can perceive the joining anomaly and state in real time or analyze and summarize the historical joining data.The practical application shows that the system can accurately and efficiently perceive and analyze the process of fuselage join.