International Journal of Reliability, Quality and Safety Engineering

In the past 35 years numerous software reliability growth models (SRGMs) have been proposed under diverse testing and debugging (T&D) environments and applied successfully in many real life software projects but no SRGM can claim to be the best in general as the physical interpretation of the T&D is not general. Unified modeling approach proves to be very successful in this regard and provides an excellent platform for obtaining several existing SRGM following single methodology. It forms the main focus of this paper; here we propose a unification modeling approach applying the infinite server queuing theory based on the basic assumptions of an SRGM defining three level of complexity of faults. Our unification methodology can be used to obtain several existing and new SRGMs which consider testing a one stage process with no fault categorization, two/three stage process with random delay functions and hence categorize faults in two/three level complexity. We have also provided data analysis based on two actual T&D data set for some of the models discussed and proposed in the paper.

Classification, an important data mining function that assigns class label to items in a collection, is of practical applications in various domains. In software engineering, for instance, a common classification problem is to determine the quality of a software item. In such a problem, software metrics represent the independent features while the fault proneness represents the class label. With many classification problems, one must often deal with the presence of irrelevant features in the feature space. That, coupled with class imbalance, renders the task of discriminating one class from another rather difficult. In this study, we empirically evaluate our proposed wrapper-based feature ranking where nine performance metrics aided by a particular learner and a methodology are considered. We examine five learners and take three different approaches, each in conjunction with one of three different methodologies: 3-fold Cross-Validation, 3-fold Cross-Validation Risk Impact, and a combination of the two. In this study, we consider two sets of software engineering datasets. To evaluate the classifier performance after feature selection has been applied, we use Area Under Receiver Operating Characteristic curve as the performance evaluator. We investigate the performance of feature selection as we vary the three factors that form the foundation of the wrapper-based feature ranking. We show that the performance is conditioned by not only the choice of methodology but also the learner. We also evaluate the effect of sampling on wrapper-based feature ranking. Finally, we provide guidance as to which software metrics are relevant in software defect prediction problems and how the number of software metrics can be selected when using wrapper-based feature ranking.

Predicting the reliability of software systems based on a component approach is inherently difficult, in particular due to failure dependencies between the software components. Since it is practically difficult to include all component dependencies in a system's reliability calculation, a more viable approach would be to include only those dependencies that have a significant impact on the assessed system reliability. This paper starts out by defining two new concepts: data-serial and data-parallel components. Then, this paper illustrates how the components' marginal reliabilities put direct restrictions on the components' conditional probabilities, and proves that the degrees of freedom are much fewer than first anticipated when it comes to conditional probabilities. At last, a test system, consisting of five components, is investigated to identify possible rules for selecting the most important component dependencies. To do this, three different techniques are applied: (1) direct calculation, (2) Birnbaum's measure and (3) Principal Component Analysis (PCA). The results from the analyses clearly show that including partial dependency information may give substantial improvements in the reliability predictions, compared to assuming independence between all software components. The analyses also indicate that including only dependencies between data-parallel components may give predictions close to the system's true failure probability, as long as the dependency between the most unreliable components is included. Including only dependencies between data-serial components may however result in predictions even worse than by assuming independence between all software components.

In industry, bogey testing, also known as the zero-failure testing, is often used to demonstrate that a product achieves the required reliability at a high confidence level. This test method is simple to apply; however, it requires excessive test time and/or a large sample size, and thus is usually unaffordable. For some products whose failure is defined as a performance characteristic exceeding a threshold, it is possible to measure the performance characteristic during testing. The measurement data can be employed to predict whether or not a test unit will fail by the end of test. When there are sufficient data to make such a prediction with a high degree of confidence, the test of the unit can be terminated. As a result, the test time is reduced. Yang studies the test time reduction for Weibull and binomial distributions.¹ This paper describes the test method for lognormal distribution. In particular, this paper describes the sample size, degradation models, and cost function for the lognormal distribution. Then the paper describes the optimum test plans, which choose the optimal sample size and the expected test time, by minimizing the total test cost and simultaneously satisfying the constraints on the type II error and the available sample size. An example is given to illustrate the test method.

Modeling correlated component failures poses a unique challenge for reliability researchers because it requires ingenuity to devise an approach free from the assumption that components fail in a statistically independent manner. Several studies have addressed this problem with models that introduce additional parameters to describe the correlated failure of components. However, these earlier techniques often require the correlations to be positive and almost always introduce an exponential number of correlation parameters. These restrictions limit the scalability of existing approaches for conducting sensitivity analysis on the correlation parameters, which could identify correlation reductions that would improve system reliability. This paper presents a technique for reliability and sensitivity analysis that requires only a quadratic number of correlation parameters, encompassing systems with both negative and positive component correlations. Unlike previous research, the proposed approach places no unnecessary restrictions on a system's correlation parameters. A series of examples illustrates the flexibility of the approach. The results quantitatively confirm that negative component correlation assists fault-tolerant systems to attain levels of reliability even higher than systems of statistically independent redundant components. Thus, the techniques introduced here offer a methodology to concisely measure the utility of negative component correlations on system reliability improvement.