Comparative study on logical analysis of data (LAD), artificial neural networks (ANN), and proportional hazards model (PHM) for maintenance prognostics

Abstract

PurposeCondition-based maintenance (CBM) has become a central maintenance approach because it performs more efficient diagnoses and prognoses based on equipment health condition compared to time-based methods. CBM models greatly inform maintenance decisions. This research examines three CBM fault prognostics models: logical analysis of data (LAD), artificial neural networks (ANNs) and proportional hazard models (PHM). A methodology, which involves data pre-processing, formulating the models and analyzing model outputs, is developed to apply and compare these models. The methodology is applied on NASA’s Turbofan Engine Degradation data set and the structural health monitoring (SHM) data set from a Nova Scotia Bridge. Results are evaluated using three metrics: error, half-life error and a cost score. This paper concludes that the LAD and feedforward ANN models compares favorably to the PHM model. However, the feedback ANN does not compare favorably, and its predictions show much larger variance than the predictions from the other three methods. Based on these conclusions, the purpose of this paper is to provide recommendations on the appropriate situations in which to apply these three prognostics models.Design/methodology/approachLAD, ANNs and PHM methods are adopted to perform prognostics and to calculate the mean residual life (MRL) of eqipment using NASA’s Turbofan Engine Degradation data set and the SHM data set from a Nova Scotia Bridge. Statistical testing was used to evaluate the statistical differences between the approaches based on these metrics. By considering the differences in these metrics between the models, it was possible to draw conclusions about how the models perform in specific cases.FindingsResults were evaluated using three metrics: error, half-life error and a cost score. It was concluded that the LAD and feedforward ANN models compares favorably to the PHM model. However, the feedback ANN does not compare favorably and its predictions show much larger variance than the predictions from the other three methods. Overall the models predict failure after it has already occurred (negative error) when the residual life is large and vice versa.Practical implicationsIt was concluded that a good CBM prognostics model for practical implications can be determined based on three main considerations: accuracy, run time and data type. When accuracy is a main concern, as in the case where impacts of failure are large, LAD and feedforward neural network are preferred. The preference changes when run time is considered. If data can be easily collected and updating the model is performed often, the ANNs and LAD are preferred. On the other hand, if CM data are not easily obtainable and existing data are not representative of the population’s behavior, data type comes into play. In this case, PHM is preferred.Originality/valuePrevious research in the literature performed reviews of multiple independent studies on CBM techniques performed on different data sets. They concluded that it is typically harder to implement artificial intelligence models, because of difficulties in data procurement, but these approaches offer improved performance as compared to more traditional model-based and statistical approaches. In this research, the authors further investigate and compare the performance and results from two major artificial intelligence models, namely, ANNs and LAD, and one pioneer statistical model, PHM over the same two real life prognostics data sets. Such in-depth comparison and review of major CBM techniques was missing in current literature of CBM field.