Process monitoring, diagnostics and prognostics using support vector machines and hidden Markov models
Condition-Based Maintenance (CBM) technology increases system availability and safety while reducing costs, attributed to reduced maintenance and inventory, increased capacity, and enhanced logistics and supply chain performance. Employing effective generic process monitoring methods for abrupt failures and diagnostics and prognostics algorithms for incipient failures is an important prerequisite for widespread deployment of CBM.
Diagnostics is the process of identifying, localizing and determining severity of a machine failure, whereas prognostics is the process of estimating the remaining-useful-life (RUL). In contrast to prognostics, there exist many methods in the diagnostics literature. However, most generic diagnostic algorithms cannot effectively detect failure modes in a timely manner. A generic, machine independent method for diagnostics and prognostics is the dream of the researcher's in this field and the focus of this research.
This work presents methods based on support vector machines and hidden Markov models to diagnose abrupt and incipient failures and to estimate the RUL. The presented methods have the ability to handle non-stationary processes. There exist three major goals for this dissertation: detecting abrupt failures (i.e. process monitoring), identifying the state of incipient failures in advance (i.e. health state estimation) and estimating RUL of the machine (i.e. prognostics).
A General Support Vector Representation Machine (GSVRM) based on novelty detection principles is proposed for process monitoring. GSVRM is a non-parametric method and does not make strong assumptions about auto-correlation structure of the data. In addition, it requires only ‘normal’ data for training. However, it can learn from failure data when available. GSVRM is implemented on benchmarking datasets in the literature as well as synthetic datasets.
Variants of hidden Markov models (i.e. regular, auto-regressive, and hierarchical HMM) are implemented for health state estimation and prognostics. Implementation of HMM as dynamic Bayesian network dramatically reduces the number of parameters and gives us more flexibility in model structure design. In prognostics, Monte-Carlo simulation using Markov models is employed to estimate RUL distribution. The proposed health state estimation and prognostics methods are applied to a drilling process, the most popular industrial machining process. The results of all three modules are very promising and reported in the dissertation.