Nikita Birbasov

(Advisor: Prof. Dimitri Mavris)

will propose a doctoral thesis entitled,

A METHOD FOR DIAGNOSTICS AND PROGNOSTICS WITH ADAPTATION

FOR PREDICTION OF NOVEL FAULTS 

On

Wednesday, April 16 at 2:00 p.m. 
Weber Space and Technology Building (SST II)
Collaborative Visualization Environment (CoVE)  

Teams Meeting: Microsoft Teams

Abstract
Spacecraft failure is a known and expensive problem that has plagued national and private entities since the launch of Sputnik. Some failures lead to a tragic loss of crew, while others lead to the loss of millions of dollars in investment. To address this, the aerospace industry, along with other industries, has been implementing varying degrees of fault management. This comes in the form of hardware redundancy or software techniques. Hardware redundancy involves installing copies of hardware that work alongside each other. This type of approach has a detrimental impact on system cost and weight. Software techniques are a part of the growing field of systems designed for fault diagnosis and prognosis as well as possible resolution. The current drive is to enable condition-based maintenance, where software systems can detect faults, isolate them to the component level, and then predict remaining useful life and preemptively request maintenance before the component fails. The advancement of these techniques is important to the future of spacecraft travel as humanity tries to reach further away from Earth. 

Currently planned missions will set up habitats around the moon and possibly send crewed missions to Mars; all of this involves long travel through a very hostile environment. The further away a spacecraft is, the longer it takes for communication and the harder it becomes to resolve faults through communications with ground crews. This is a limitation imposed by two factors: the communication times and the transmission bandwidth. Simply, the sensor information on board of the spacecraft is going to be much richer than anything it will send back to Earth. Many of these new crafts are meant to only be crewed for very short periods of time (e.g. Lunar Gateway), thus it is important that it is kept operational autonomously. This drives the need to create diagnostic systems capable of identifying and resolving faults with higher autonomy. The current prognostic health management (PHM) frameworks are implemented with the use of physics-based modeling, machine learning, and statistical analysis. Yet many of them are limited to specific faults and components, with very limited capabilities to learn and adapt to unexpected faults.

This work aims to expand the capabilities of diagnostic and prognostic systems to recognize and predict the impact of known and unknown faults and introduce the ability to learn fault signatures and effects as the fault develops. To do this, the work will introduce a methodology that follows the general outline of diagnostic and prognostic systems with the addition of case-based reasoning and online machine learning for the purposes of online diagnostics and prognostics. All the components will be designed to interact with time series data as provided by sensors at run time. Within the proposed methodology, an initial database of system operation and the most likely faults would need to be constructed and used for training to create the initial set of cases. The proposed diagnostic phase will consist of anomaly detection and diagnostic logic. Anomaly detection will determine if the behavior is deviating from a healthy mode of operation. The anomaly detection will be carried out with matrix profile and extraction of representative primitives. Diagnostic logic will be divided into two steps. First is the classification of the anomaly as part of a known fault class or as a novel occurrence, which will be done with open set recognition classification. Second, if the anomaly is novel, Bayes logic will be used to identify if it is a possible signature of a fault or some other occurrence. If the system recognizes a fault, it will extract the case out of the case base. As this work is limited to diagnosis and prognosis, the information within the case base will be a model capable of predicting system operation into the future. These models will be constructed with the use of supervised machine learning regression techniques. In the case of an unknown fault, that system will enter an online learning mode where a new case will be created. A known case that is closest to the currently occurring one will be used to enable transfer learning by doing parameter transfer and online tuning to create a model capable of predicting system remaining useful life under the unknown fault.

The methodology will then be tested on a data set from an ion mill etch tool . The data set consists of multiple time-series sensor outputs as well as system operational inputs. The data set has three different fault modes. While the data set is not that of a spacecraft, it does provide a set of time series sensor information. As the type of data provided in the data set is similar and the overall diagnostic system is meant to be flexible, it was judged by the author to be an appropriate data set for testing. The final demonstration will involve application of the designed system to this data set while withholding at least one fault from training and showing the capability of the system to identify and predict known faults and the ability to learn a newly occurring trend at runtime.

Committee

• Prof. Dimitri Mavris – School of Aerospace Engineering (Advisor)
• Prof. Kyriakos Vamvoudakis – School of Aerospace Engineering
• Prof. Glen Chou– School of Aerospace Engineering
• Dr. Michael Balchanos – School of Aerospace Engineering
• Dr. Stephen Edwards – NASA Marshall Space Flight Center