Prognostic performance is associated with accurately estimating remaining useful life. Difficulty in accurate prognostic applications can be tackled by processing raw sensor readings into more meaningful and comprehensive health condition indicators that will then provide performance information for remaining useful life estimations. To that end, typically, multiple tasks on data pre-processing and predictions have to be carried out such that tasks can be assessed using different methodological aspects. However, incompatible methods may result in poor performance and consequently lead to undesirable error rates.

The present research evaluates data training and prediction stages. A data-driven prognostic method based on a feed-forward neural network framework is first defined to calculate the performance of a complex system. Then, the health indicators are used in a similarity based remaining useful life estimation method. This framework presents a conceptual prognostic protocol that overcomes challenges presented by multi-regime condition monitoring data.