Abstract
The ability to detect when a system undergoes an incipient fault is ofparamount importance in preventing a critical failure. In this work, we proposean information-driven fault detection method based on a novel concept driftdetector. The method is tailored to identifying drifts in input-outputrelationships of additive noise models (i.e., model drifts) and is based on adistribution-free mutual information (MI) estimator. Our scheme does notrequire prior faulty examples and can be applied distribution-free over a largeclass of system models. Our core contributions are twofold. First, wedemonstrate the connection between fault detection, model drift detection, andtesting independence between two random variables. Second, we prove severaltheoretical properties of the proposed MI-based fault detection scheme: (i)strong consistency, (ii) exponentially fast detection of the non-faulty case,and (iii) control of both significance levels and power of the test. Toconclude, we validate our theory with synthetic data and the benchmark datasetN-CMAPSS of aircraft turbofan engines. These empirical results support theusefulness of our methodology in many practical and realistic settings, and thetheoretical results show performance guarantees that other methods cannotoffer.