Unsupervised Learning of Data Representations and Cluster Structures: Applications to Large-scale Health Monitoring of Turbofan Aircraft Engines

Overview & Synopsis

This thesis is interested in unsupervised statistical learning methods and their applications to health monitoring of aircraft engines at an industrial scale. Our first objective is to make health monitoring methodologies scale to massive datasets and allow engineering team to flexibly deploy various use cases. Besides the engineering aspects, we also try to address two fundamental theoretical challenges in unsupervised learning. First, the links between cluster structure and representation.And second, the very definition of structure, arising from the problem of model selection in clustering.Modern aircraft engines generate growing amounts of data during manufacturing,tests and flights, that can be leveraged for health monitoring and predictive maintenance, in order to improve safety, availability and reduce costs. In this work, we use sensor measurements collected on board of civil short and mid-range aircraft.These data sets are temporal and highly multidimensional due to the large number of sensors and sampling frequencies. Hence, applications need to scale to the large volumes of data, driven by the growing number of daily operating engines.Among all unsupervised learning approaches, clustering and self-organizing maps (SOM) provide useful insights on the distribution of complex and high-dimensional unlabeled data sets. To describe the internal state of an engine, expert indicators or features need to be extracted from raw data, before applying clustering algorithms.Our first contribution is to scale these methodologies using Big Data tools and distributed computing, in order to process entire fleets. We propose a generic and scalable pipeline enabling engineers to analyze flight data on a cluster. In addition,we present an new application to monitoring of vibration signatures.Another option is to automatically extract relevant features with deep neural networks, known as deep learning, which had a great impact in many areas of machine learning. Recently, its ability to improve clustering has been investigated. A second contribution of this thesis is a Deep Embedded SOM, a neural network-based model performing joint representation learning with an autoencoder and self-organization of the cluster prototypes. The third contribution of this thesis concerns model selection, which is perhaps the most difficult problem in clustering. We adopt the framework of cluster stability analysis, and propose a novel concept of within-cluster stability, leading to a criterion called Stadion (stability difference criterion) able to effectively select the number of clusters in a data set. We also apply Stadion to time series clustering validation by leveraging invariant transformations of the data.