Fault-tolerant sensor fusion for aircraft height estimation

Wiegman, Adrian Peter (2019-04)

Thesis (MEng)--Stellenbosch University, 2019.

Thesis

ENGLISH ABSTRACT: This thesis presents the design and verification of a fault-tolerant sensor fusion system to provide robust height estimates for commercial airliners. Current commercial aircraft systems obtain a height estimate by taking the median of redundant radio altimeter measurements. However, this approach is not robust to a failure that affects all radio altimeters. The proposed system uses technology redundancy by combining the measurements from the available aircraft sensors that provide either height or altitude measurements: the radio altimeter, the GPS, the inertial sensors, and the instrument landing system (ILS). An onboard terrain map is used to convert altitude measurements to height measurements. The robust height estimation problem is divided into two major sub-problems: fault detection and isolation and sensor fusion. To address the fault detection and isolation sub-problem, a variety of data-driven and model-based techniques were investigated. Two general approaches were considered: fault diagnosis using sensor measurements from a single time instant, and fault diagnosis using sensor measurements from a window of consecutive time instants. For the single time instant approach, only data-driven techniques were applied, including outlier detectors, binary classifiers, and multi-class classifiers. For the multiple time instant approach, both model-based and data-driven techniques were applied. The model-based techniques included the Bank of Kalman filters and the Robust Kalman filter, while the data-driven techniques include Model Consensus, Dynamic Principal Component Analysis, and Gaussian Naïve Bayes with time as an additional variable. Finally, to address the sensor fusion sub-problem, three methods were investigated: using the median of the sensor measurements, using the weighted average of the sensor measurements, and using a Kalman filter to perform optimal sensor fusion. A simulation model was used to generate synthetic training and testing data for the data-driven techniques. Mathematical models were established for the aircraft motion, the sensors, and the terrain. The structure and nominal parameters for the sensor models were based on information sourced from literature, and then the sensor parameters were tuned to fit a real dataset provided by Airbus. Fault models for six types of sensor faults were also created. The simulation model was used to generate a large dataset of representative sensor measurements containing both \no-fault" and \fault" conditions. The fault detection and isolation, and the sensor fusion were tested using both simulated data and real datasets of actual ight data with synthetic sensor failures injected. The single time instant approach achieved fault diagnosis accuracies from 85% (Support Vector Machine) to 94% (k-Nearest Neighbors). The multiple time instant approach achieved fault diagnosis accuracies from 93% (Model Consensus) to 99% (Robust Kalman filter). The three sensor fusion approaches produced height estimates with average accuracies from 5:1m to 7:4 m.

AFRIKAANSE OPSOMMING: Hierdie tesis beskryf die ontwerp en verifikasie aan van 'n fouttolerante sensorfusie-stelsel om betroubare hoogte afskattings vir kommersiële vliegtuie te verskaf. Huidige kommersiële vliegtuigstelsels verkry 'n hoogte afskatting deur die mediaan van veelvuldige radio-hoogtemeter metings te neem. Hierdie benadering is egter nie betroubaar vir 'n fout wat alle radio-hoogtemeters gelyktydig beïnvloed nie. Die voorgestelde stelsel gebruik tegnologie oorbodigheid deur die metings te kombineer van die beskikbare vliegtuig sensors wat hoogte bo grond of hoogte bo seevlak meet: die radio-hoogtemeter, die GPS, die inersiële sensors, en die instrumentlandingstelsel (ILS). Die robuuste hoogte afskatting probleem word verdeel in twee subprobleme: foutopsporing en isolasie, en sensorfusie. Om die foutopsporing en isolasie subprobleem aan te spreek, is 'n verskeidenheid data-gedrewe en modelgebaseerde tegnieke ondersoek. Twee algemene benaderings is oorweeg: foutdiagnose met behulp van sensormetings vanaf 'n enkele tydstip, en foutdiagnose met behulp van sensormetings vanaf 'n venster van opeenvolgende tydstippe. Vir die enkel tydstip benadering is verskeie data-gedrewe tegnieke toegepas, insluitende uitskieter opspoorders, binêre klassifiseerdes en multi-klas klassifiseerders. Vir die meervoudige tydstip benadering is beide modelgebaseerde en data-gedrewe tegnieke toegepas. Die modelgebaseerde tegnieke het die Bank van Kalman filters en die Robuuste Kalman filter ingesluit, terwyl die data-gedrewe tegnieke Model Konsensus en Dinamiese Hoofkomponent Analise. Laastens, om die sensorfusie subprobleem aan te spreek is drie metodes ondersoek: gebruik van die mediaan van die sensormetings, gebruik van die geweegde gemiddelde van die sensormetings, en gebruik van 'n Kalman filter vir optimaal sensorfusie. 'n Simulasiemodel is ontwikkel om sintetiese opleiding en toetsingdata te skep vir die data-gedrewe tegnieke. Wiskundige modelle is afgelei vir die vliegtuig beweging, die sensors, en die terrein. Die struktuur en nominale parameters vir die sensormodelle is gegrond op inligting wat uit literatuur verkry is, waarna die sensorparameters ingestel is om 'n werklike datastel, wat deur Airbus verskaf is, te pas. Foutmodelle vir ses tipes sensorfoute is ook geskep. Die simulasiemodel is gebruik om 'n groot datastel van verteenwoordigende sensormetings te genereer wat beide \geen fout" en \fout" toestande bevat. Die foutopsporing en isolasie en die sensorfusie is getoets met behulp van beide gesimuleerde data en datastelle van werklike vlugdata met byvoeging van sintetiese sensorfoute. Die enkel tydstip benadering foutdiagnose akkuraathede behaal van 85% (\Support Vector Machine") tot 94% (k-Naaste Bure). Die meervoudige tydstip benadering het foutdiagnose akkuraathede behaal van 93% (Model Konsensus) tot 99% (Robuuste Kalman filter). Die drie sensor-fusie benaderings het hoogte afskattings met gemiddelde akkuraatheid van 5:1m tot 7:4m gelewer.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/105895
This item appears in the following collections: