International Journal of Reliability, Quality and Safety Engineering

A linear consecutive $k$-out-of-$n$:failed system, henceforth written consecutive $k$-out-of-$n$:F, is an $n$-component system that fails if $k$ or more consecutive components in positions $i,(i+1),\dots ,(i+k-1)$ fail, where $1\le i\le n-(k-1)$. The consecutive $k$-out-of-$n$:F system model is particularly valuable for characterizing critical infrastructures such as telecommunications that provide essential services to society. However, events that impact critical infrastructures such as natural disasters often produce geospatially correlated failures, necessitating reliability models that can accommodate these scenarios where component failures may be correlated.

This paper presents a method to explicitly model the correlation between component failures of a consecutive $2$-out-of-$n$:F system. The approach produces analytical expressions for the reliability of discrete and continuous systems in terms of component reliabilities and correlation between the component failures. The explicit correlation parameter simplifies sensitivity analysis for a variety of measures of interest, including reliability, density function, hazard rate, mean time to failure (MTTF), availability, and mean residual life (MRL). It is illustrated through examples where correlated failures can negatively influence system reliability. Thus, the approach can quantify improvements to system reliability that could be achieved by lowering the correlation between failures of system components.

As the technology has developed, cloud computing which is a kind of Internet-based computing has widely spread. In terms of operating cloud system in remote locations, there exist some problems when the failure of cloud system occurs. In such cases, workers need to go to the remote locations and maintain the system. This paper considers stochastic models of a redundant cloud system in remote locations. The cloud system has active devices and standby devices. We derive the steady-state availability and discuss optimal policies which maximize it. Finally, numerical examples are given.

Condition-based maintenance (CBM) optimization involves considering inherent uncertainties and external uncertainties. Since computational complexity increases exponentially with the number of degradation uncertainties and stages, scenario reduction aims to select small set of typical scenarios which can maintain the probability distributions of outputs of possible scenarios. A novel scenario reduction method, 3D-outputs-clustering scenario reduction (3DOCS), is presented by considering the impacts of uncertainty parameters on the output performance for CBM optimization which have been overlooked. Since the output performance for CBM is much more essential than the inputs, the proposed scenario reduction method reduces degradation scenarios by $K$-means clustering of the multiple outputs of degradations scenarios for CBM. It minimizes the probabilistic distribution distances of outputs between original and selected scenarios. Case studies show that 3DOCS has advantages as a smaller distance of output performance of selected scenarios compared to that of initial scenarios.

In this paper, an artificial neural network (ANN)-based dynamic weighted combination model trained by novel particle swarm optimization (PSO) algorithm is proposed for software reliability prediction. Different software reliability growth models (SRGMs) are merged based on the weights derived by the learning algorithm of the proposed ANN. To avoid trapping in local minima during training of the ANN, we propose a neighborhood-based adaptive PSO (NAPSO) algorithm for learning of the proposed ANN in order to find global optimal weights. We conduct the experiments on real software failure data sets for validation of the proposed dynamic weighted combination model (PDWCM). Fitting performance and predictability of the proposed PSO-based neural network are compared with the conventional PSO-based neural network (CPSO) and existing ANN-based software reliability models. We also compare the performance of the proposed PSO algorithm with the CPSO algorithm through learning of the proposed ANN. Empirical results indicate that the proposed PSO and CPSO-based neural network present fairly accurate fitting and prediction capability than the other existing ANN-based software reliability models. Moreover, the proposed PSO-based neural network is most promising for the purpose of software fault prediction since it shows comparatively better fitting and prediction performance results than the other models.

Non-homogeneous Poisson process (NHPP) software reliability growth models (SRGM) enable quantitative metrics to guide decisions during the software engineering life cycle, including test resource allocation and release planning. However, many SRGM possess complex mathematical forms that make them difficult to apply. Specifically, traditional procedures solve a system of nonlinear equations to identify the numerical parameters that best characterize failure data. Recently, researchers have developed expectation-maximization (EM) algorithms for NHPP SRGM that exhibit better convergence properties and can therefore find maximum likelihood estimates with greater ease.

This paper presents an adaptive EM (AEM) algorithm, which combines an earlier EM algorithm for NHPP SRGM with unconstrained search of the model parameter space. Our performance analysis shows that the AEM outperforms state-of-the-art EM algorithms for NHPP SRGM with very strong statistical significance, which is as much as hundreds of times faster on some data sets. Thus, the approach can fit SRGM very quickly. We also incorporate this high performance adaptive EM algorithm into a heuristic nested model selection procedure to objectively select a model of least complexity that best characterizes the failure data. Results indicate this heuristic approach often identifies the model possessing the best model selection criteria.