Table 3.

Advantages and disadvantages of specific prognostic and health-management techniques.

	Pros.	Cons.	Application
Bayesian Network	easy to understand and transparent encodes expert knowledge used for both RUL and RCA purposes handles uncertainty	complex preparation process both expert and analytical knowledge required finds only known/defined cases	fault detection [158] diagnosis [194,195] scheduling [196] RUL [197,198]
SVM	good modeling of non-linear and linear relationships used for both regression and classification does not require a large learning set	lack of transparency data scientist’s knowledge needed with large datasets, long computation times	fault detection [137,138,185] condition monitoring [199,200]
PCA	handles multidimensional datasets works well with other techniques generalizes the data	loss of some information features lose linking to specific components	fault detection [131,132,201,202]
Expert system	transparent and easy to understand good interaction with domain knowledge no need for a physical process model	advanced models require a strong effort works only with defined cases	fault detection, planning [203] fault detection [204,205,206]
Fuzzy logic	extends the capabilities of the expert system to time series analysis deals with input noise and uncertainty	requires knowledge to apply fuzzy rules	fault detection [122] diagnostics [207]
Physical models	provide precise results for a specific well-known case/process algorithms understandable for industry experts	require considerable modeling effort and extensive domain knowledge	RUL [208] condition monitoring [209]
ANNs	provide the capability to model complex, non-linear relationships no domain knowledge required can be used in conjunction with other techniques provide direct result output	“black box” results may be non-transparent prone to overfitting difficult in determining the uncertainty of results require a large training set	RUL [178,179]
ARIMA	computationally efficient does not require large datasets requires no expert knowledge	short term forecast only sensitive to noise and process variations	RUL [210,211,212,213,214]
HMMs	allow modeling of both time series and stationary data handle incomplete datasets	computationally complex do not detect previously undefined events	fault detection [126,127] RUL [215,216]