Self-monitoring solutions first appeared to avoid catastrophic breakdowns in safety-critical mechanisms. The design behind these solutions relied heavily on the physical knowledge of the mechanism and its fault. They usually involved installing specialized sensors to monitor the state of the mechanism and statistical modeling of the recorded data. Mainly, these solutions focused on specific components of a machine and rarely considered more than one type of fault.

In our work, on the other hand, we focus on self-monitoring of complex machines, systems composed of multiple components performing heterogeneous tasks and interacting with each other: systems with many possible faults. Today, the data available to monitor these machines is vast but usually lacks the design and specificity to monitor each possible fault in the system accurately. Some faults will show distinctive symptoms in the data; some faults will not; more interestingly, there will be groups of faults with common symptoms in the recorded data.

The thesis in this manuscript is that we can exploit the similarities between faults to train machine learning models that can significantly improve the performance of self-monitoring solutions for complex systems that overlook these similarities. We choose to encode these similarity relationships into hierarchies of faults, which we use to train hierarchical supervised models. We use both real-life problems and standard benchmarks to prove the adequacy of our approach on tasks like fault diagnosis and fault prediction.

We also demonstrate that models trained on different hierarchies result in significantly different performances. We analyze what makes a good hierarchy and what are the best practices to develop methods to extract hierarchies of classes from the data. We advance the state-of-the-art by defining the concept of heterogeneity of decision boundaries and studying how it affects the performance of different class decompositions.