As systems and equipment become more complex and expensive, the competition drives industries to become increasingly concerned about system reliability and availability. The need to reduce maintenance and logistics costs while at the same time minimizing the risk of catastrophic failures, and maximizing system availability is leading a drive towards predictive maintenance. In predictive maintenance, activities take place only when there is objective evidence of an impending fault or
failure condition, while at the same time ensuring safety, reliability, and reducing the overall total life costs.
One important capability, and the most amongst others, of predictive maintenance is the estimation, or prognosis, of the remaining useful life (RUL) of a system. Concerning this, there are two main types of predictive maintenance categories; namely model-based methods and data-driven methods.
- Model-based methods: These approaches use constructed mathematical models to describe the physics of the system and failure modes under observation, derived from first principles. This has been used traditionally to understand component failure mode progression. The main benefit in this, is the ability to incorporate physical understanding of the process under observation, thus delivering increased accuracy and can address subtle performance problems. The main drawback, is the difficulty or infeasibility of constructing a mathematical model that incorporates the system’s behavior.
- Data-driven methods: These approaches use actual data (gathered with sensors or operator measures) to approximate and track features revealing the underlying degradation of components and to forecast the (global) behavior of the system under observation. These include machine learning and artificial intelligence techniques (neural networks, random forests, support vector machines, hidden markov models etc) and statistical techniques (multivariate methods, partial least squares, etc). The main benefit is that there is no need of explicitly understanding the underlying physics of the system and the main drawback is that these techniques are “data-hungry” and require historical fault/failure data in terms of time plots of various signals leading up to failure, to train on.
- Lastly, there is a middle solution which is a combination of model-based and data-driven methods, called hybrid approach. In these approaches techniques from both categories are used (e.g. synthetic training data derived from an accurate model to train data-driven algorithms which will be used to make predictions on sensor data).
Generally, the reduction in applicability is a reflection of the increasing complexity/cost of the different approaches. See picture below :