Earticle

소프트웨어 결함 예측(SDP)은 오류가 발생할 가능성이 있는 모듈을 사전에 식별하여 소프트웨어 개발의 효율을 높이고 있다. SDP에서의 주과제는 예측 성능을 향상시키는것에 있다. 최근 연구에서는 딥러닝 기법이 소프트웨어 결함 예측(SDP) 분야에 적용되어 있으며, 특히 구조화된 데이터를 분석하는 데 뛰어난 성능을 보이고 있는 SAINT 모델이 주목받고 있다. 본 연구는 SAINT 모델을 다른 주요 모델(XGBoost, Random Forest, CatBoost)과 비교하여 SDP에 적용 가능한 최신 딥러닝 기법을 조사하였다. SAINT는 일관되게 우수한 성능을 보여주며 결함 예측 정확도 향상에 효과적임을 입증하였다. 이 연구 결과는 실용적인 소프트웨어 개발 상황에서 결함 예측 방법론을 발전시킬 수 있는 SAINT의 잠재력을 강조하며, 교차 검증, 특성 스케일링, 비교 분석 등을 포함한 철저한 방법론을 통해 수행되었다.

Software Defect Prediction (SDP) enhances the efficiency of software development by proactively identifying modules likely to contain errors. A major challenge in SDP is improving prediction performance. Recent research has applied deep learning techniques to the field of SDP, with the SAINT model particularly gaining attention for its outstanding performance in analyzing structured data. This study compares the SAINT model with other leading models (XGBoost, Random Forest, CatBoost) and investigates the latest deep learning techniques applicable to SDP. SAINT consistently demonstrated superior performance, proving effective in improving defect prediction accuracy. These findings highlight the potential of the SAINT model to advance defect prediction methodologies in practical software development scenarios, and were achieved through a rigorous methodology including cross-validation, feature scaling, and comparative analysis.

Litigation associated with construction defects and quality issues in apartment housing has markedly increased in Korea. A key objective of such litigation is to estimate defect repair costs, an issue of mutual significance for both residents and project developers. This study develops a regression-based predictive model to estimate defect repair costs in apartment complexes. Based on an integrated review of prior studies, construction cost, its natural logarithm, number of dwelling units, construction period, and elapsed period were selected as major explanatory variables. When predicting the defect repair cost or its logarithmic form, the proposed model satisfied the key assumptions of regression analysis independence and absence of multicollinearity while achieving the highest explanatory power. Future research should further explore additional determinants and refine analytical methodologies for modeling defect repair costs in multi-family housing.

Prediction of a minimum crack size for growth, which is defined as a crack size that grows fast enough to keep ahead of its removal by contact wear and periodic grinding, is the most demanding work to prevent rail from fatigue failure and develop cost effective railway maintenance strategy In this study, we investigated the wheel load increment due to a rail defect during a train ran over it, and its effect on the minimum crack size for growth. For this purpose, we developed simulation software based on the Fletcher and Kapoor's "2.5D" model and measured wheel load increment during a train passed over a defect. A maximum contact pressure and contact patch size were calculated by 3D FEM and crack growth analyses were performed by varying two of dominant contact contributors; surface friction coefficient(0.1, 0.2, 0.3 and 0.4) and crack aspect ratio. The minimum crack sizes for growth were calculated from 0.29 to 1.44mm depending on the contact conditions. They were decreasing with increasing surface friction coefficient and decreasing with crack aspect ratio(a/b).

Determining when software testing should begin and the number of resources that may be required in order to find and fix defects are complicated decisions. If we can predict the number of defects for an upcoming software product given the current development team, it will enable us to make better decisions. A majority of reported defects are managed and tracked using a defect life cycle, which tracks a defect throughout its lifetime. The process starts when the defect is found and ends when the resolution is verified and the defect is closed. Defects transition through different states according to the evolution of the project, which involves testing, debugging, verification. In paper, we presents defect prediction model for consecutive software products that is based on entropy.

Finding defects in a software system is not easy. Effective detection of software defects is an important activity of software development process. In this paper, we propose an approach to predict residual defects, which applies machine learning algorithms (classifiers) and defect distribution model. This approach includes two steps. Firstly, use machine learning Algorithms and Association Rules to get defect classification table, then confirm the defect distribution trend referring to several distribution models. Experiment results on a GUI project show that the approach can effectively improve the accuracy of defect prediction and be used for test planning and implementation.

Predicting the existing defects in software products is one of the considerable issues in software engineering that contributes a lot toward saving time in software production and maintenance process. In fact, finding the desirable models for predicting software defects has nowadays turned into one of the main goals of software engineers. Since intricacies and restrictions of software development are increasing and unwilling consequences such as failure and errors decrease software quality and customer satisfaction, producing error-free software is very difficult and challenging. One of the efficient models in this field is multilayer neural network with proper learning algorithm. Many of the learning algorithms suffer from extra overfitting in the learning datasets. In this article, setting multilayer neural network method was used in order to improve and increase generalization capability of learning algorithm in predicting software defects. In order to solve the existing problems, a new method is proposed by developing new learning methods based on support vector machine principles and using evolutionary algorithms. The proposed method prevents from overfitting issue and maximizes classification margin. Efficiency of the proposed algorithm has been validated against 11 machine learning models and statistical methods within 3 NASA datasets. Results reveal that the proposed algorithm provides higher accuracy and precision compared to the other models.

Software quality is a field of study and practice that describes the desirable attributes of software products. The performance must be perfect without any defects.Software quality metrics are a subset of software metrics that focus on the quality aspects of the product, process, and project.The software defectprediction model helps in early detection of defects and contributes to their efficient removal and producing a quality software system based on several metrics. The main objective of paper is to help developers identify defects based on existing software metrics using data mining techniques and thereby improve the software quality.In this paper, variousclassification techniquesare revisitedwhich are employed for software defect prediction using software metrics in the literature.

Software is a complex entity composed in various modules with varied range of defect occurrence possibility. Efficient and timely prediction of defect occurrence in software allows software project managers to effectively utilize people, cost, time for better quality assurance. The presence of defects in a software leads to a poor quality software and also responsible for the failure of a software project. Sometime it is not possible to identify the defects and fixing them at the time of development and it is required to handle such defects any time whenever they are noticed by the team members. So it is important to predict defect-prone software modules prior to deployment of software project in order to plan better maintenance strategy. Early knowledge of defect prone software module can also help to make efficient process improvement plan within justified period of time and cost. This can further lead to better software release as well as high customer satisfaction subsequently. Accurate measurement and prediction of defect is a crucial issue in any software because it is an indirect measurement and is based on several metrics. Therefore, instead of considering all the metrics, it would be more appropriate to find out a suitable set of metrics which are relevant and significant for prediction of defects in any software modules. This paper proposes a feature selection based Linear Twin Support Vector Machine (LSTSVM) model to predict defect prone software modules. F-score, a feature selection technique, is used to determine the significant metrics set which are prominently affecting the defect prediction in a software modules. The efficiency of predictive model could be enhanced with reduced metrics set obtained after feature selection and further used to identify defective modules in a given set of inputs. This paper evaluates the performance of proposed model and compares it against other existing machine learning models. The experiment has been performed on four PROMISE software engineering repository datasets. The experimental results indicate the effectiveness of the proposed feature selection based LSTSVM predictive model on the basis standard performance evaluation parameters.

The costs of finding and correcting software defects have been the most expensive activity in software development. The accurate prediction of defect‐prone software modules can help the software testing effort, reduce costs, and improve the software testing process by focusing on fault-prone module. Recently, static code attributes are used as defect predictors in software defect prediction research, since they are useful, generalizable, easy to use, and widely used. However, two common aspects of data quality that can affect performance of software defect prediction are class imbalance and noisy attributes. In this research, we propose the combination of particle swarm optimization and bagging technique for improving the accuracy of the software defect prediction. Particle swarm optimization is applied to deal with the feature selection, and bagging technique is employed to deal with the class imbalance problem. The proposed method is evaluated using the data sets from NASA metric data repository. Results have indicated that the proposed method makes an impressive improvement in prediction performance for most classifiers.

훈련 데이터 부족 문제를 해결하기 위해 제안된 교차 프로젝트 소프트웨어 결함 예측은 소스 프로젝트와 타겟 프로젝트 간의 데이터 분포 차이를 최소화하기 위해 도메인 적응 기법들을 이용한다. 본 논문은 CORAL을 사용한 새로 운 모델들을 제작하여 TCA, BDA, W-BDA를 사용한 기존 모델들과 성능을 비교하였다. 또한 딥러닝 분류기를 사용한 모델의 성능을 기존 분류기들을 사용한 모델들과 비교하였다. 평가 실험 결과 CORAL은 소수의 경우를 제외하고 대부분 의 경우 타 모델들에 비해 탁월한 성능을 보였으며, 딥러닝 분류기의 사용은 데이터 집합에 따라 모든 모델들의 성능을 좋게하거나 나빠지게하는 극단적인 결과를 보였다.

Cross-project software defect prediction has been proposed to address the issue of insufficient training data, utilizing domain adaptation techniques to minimize the distribution differences between source and target projects. This paper introduces new models using CORAL and compares their performance with existing models based on TCA, BDA, and W-BDA. Additionally, the performance of models employing deep learning classifiers is compared with those using traditional classifiers. Evaluation experiments demonstrate that, except for a few cases, CORAL generally outperforms the other models, and the use of deep learning classifiers leads to extreme outcomes, either significantly improving or deteriorating the performance of all models depending on the dataset.

매우 많은 소프트웨어 결함 예측에 관한 연구들이 수행되었으나, 학습 데이터의 부족으로 이들을 사용하기에 어려움이 있었다. 교차 프로젝트 결함 예측은 이를 해결하기 위한 기법으로 학습 데이터가 충분한 소스 프로젝트의 데이 터로 학습한 모델을 타겟 프로젝트의 결함 예측에 사용하는 것이다. 학습을 하기 전에 두 프로젝트간의 데이터 분포 차이를 최소화하기 위해 전이학습의 일종인 도메인 적응 기법들을 사용한다. 본 논문은 W-BDA, MEDA를 사용한 새로 운 모델들을 제작하여 TCA, BDA를 사용한 기존 모델들과 성능을 비교하였다. 평가 실험 결과 MEDA는 타 모델들에 비해 불규칙적이고 나쁜 성능을 보였지만 BDA는 TCA보다 더 나은 성능을 보였고, W-BDA는 BDA보다 약간 더 좋은 성능을 보였다.

Many studies on software defect prediction have been conducted, but it has been difficult to use them due to a lack of training data. Cross-project defect prediction is a technique to solve this problem, where a prediction model learned with sufficient training data from existing source project is used to predict defects in the target project. Before learning, domain adaptation techniques, a type of transfer learning, are used to minimize the difference in data distribution between the two projects. In this paper, we produced new prediction models using W-BDA and MEDA and compared their performance with existing models using TCA and BDA. As a result of the evaluation experiment, MEDA showed irregular and poor performance compared to other models, but BDA showed better performance than TCA, and W-BDA showed slightly better performance than BDA.

Software defect severity is very important in projects with limited historical data or new projects. But general software defect prediction is very difficult to collect the label information of the training set and cross-project defect prediction must have a lot of data. In this paper, an unclassified data set with defect severity is clustered according to the distribution ratio. And defect severity-based prediction model is proposed by way of labeling. Proposed model is applied CLAMI in JM1, PC4 with the least ambiguity of defect severity-based NASA dataset. And it is evaluated the value of ACC compared to original data. In this study experiment result, proposed model is improved JM1 0.15 (15%), PC4 0.12(12%) than existing defect severity-based prediction models.

Purpose: We investigated whether C-reactive protein (CRP) levels, urine protein-creatinine ratio (uProt/Cr), and urine electrolytescan be useful for discriminating acute pyelonephritis (APN) from other febrile illnesses or the presence of a cortical defect on 99mTc dimercaptosuccinic acid (DMSA) scanning (true APN) from its absence in infants with febrile urinary tract infection (UTI). Materials and Methods: We examined 150 infants experiencing their first febrile UTI and 100 controls with other febrile illnesses consecutively admitted to our hospital from January 2010 to December 2012. Blood (CRP, electrolytes, Cr) and urine tests [uProt/Cr, electrolytes, and sodium-potassium ratio (uNa/K)] were performed upon admission. All infants with UTI underwent DMSA scans during admission. All data were compared between infants with UTI and controls and between infants with or without a cortical defect on DMSA scans. Using multiple logistic regression analysis, the ability of the parameters to predict true APN was analyzed. Results: CRP levels and uProt/Cr were significantly higher in infants with true APN than in controls. uNa levels and uNa/K were significantly lower in infants with true APN than in controls. CRP levels and uNa/K were relevant factors for predicting true APN. The method using CRP levels, u-Prot/Cr, u-Na levels, and uNa/K had a sensitivity of 94%, specificity of 65%, positive predictive value of 60%, and negative predictive value of 95% for predicting true APN. Conclusion: We conclude that these parameters are useful for discriminating APN from other febrile illnesses or discriminating true APN in infants with febrile UTI.

A 3D model based on the finite element method (FEM) was built to simulate the infrared thermography (IRT) inspection process. Thermal contrast is an important parameter in IRT and was proven to be a function of defect parameters. Parametric studies were conducted on internal defects with different depths, thicknesses, and orientations. Thermal contrast evolution profiles with respect to the time of the defect and host material were obtained through numerical simulation. The thermal contrast decreased with defect depth and slightly increased with defect thickness. Different orientations of thin defects were detected with IRT, but doing so for thick defects was difficult. These thermal contrast variations with the defect depth, thickness, and orientation can help in optimizing the experimental process and interpretation of data from IRT.

소프트웨어 결함 예측은 소프트웨어 시스템 중 결함 존재할 만한 모듈을 사전에 예측하는 연구이다. 최근의 오픈 소스 프로젝트 및 공개적으로 이용 가능한 결함 정보 저장소의 증가로 인해 결함 예측 모형의 구축 및 구축된 모형의 범용성을 확인하기 위한 연구가 연구자들의 관심을 끌고 있다. 그러나 예측 모형을 새롭게 구축하는 것은 모형 구축 기술에 대한 깊이 있는 지식을 가지고 있어야 할 뿐만 아니라, 모형 구축을 위해 많은 시간이 필요한 작업이다. 모형 구축의 이러한 단점은 현실의 프로젝트에 교차 프로젝트 결함 예측 연구의 적용을 어렵게 한다. 본 논문에서는 결함 모형을 새롭게 구축하는 대신, 기 구축된 결함 예측 모형의 범용성을 확인하기 위한 연구를 수행하였다. 이를 위해 네 가지 기 구축된 결함 예측 모형을 이용하여 34가지 공개 프로젝트의 결함을 예측하는 실험을 수행하였다. 실험 결과에 따르면 기 구축된 예측 모형을 실험 대상 프로젝트의 결함 밀도에 따라서 범용성이 달라지는 것을 확인할 수 있었다. 즉, 결함 밀도가 높은 프로젝트는 상대적으로 높은 결함 예측력을 보였으며, 결함 밀도가 낮은 프로섹트는 상대적으로 낮은 결함 예측력을 보였다.

Defect prediction studies aim to predict defect-prone classes. The studies have recently focused on validating generality of prediction models because of increasing public defect information about open source projects. But constructing new prediction models is not an easy task because model developers should have specific knowledge and construction models are a time-intensive task. Such disadvantages make cross-project defect prediction study difficult for applying in the reality. In this paper, instead of constructing new prediction models, we study on validating generality of conventional prediction models. We conduct defect prediction experiments on 34 data sets using four conventional models. The results indicate that the generality of conventional models depends on defect density of a target system. That is, the models relatively predict well on high defect density projects and vice versa.

Purpose: The improvement of yield and quality in product manufacturing is crucial from the perspective of process management. Controlling key variables within the process is essential for enhancing the quality of the produced items. In this study, we aim to identify key variables influencing product defects and facilitate quality enhancement in CNC machining process using SHAP(SHapley Additive exPlanations) Methods: Firstly, we conduct model training using boosting algorithm-based models such as AdaBoost, GBM, XGBoost, LightGBM, and CatBoost. The CNC machining process data is divided into training data and test data at a ratio 9:1 for model training and test experiments. Subsequently, we select a model with excellent Accuracy and F1-score performance and apply SHAP to extract variables influencing defects in the CNC machining process. Results: By comparing the performances of different models, the selected CatBoost model demonstrated an Accuracy of 97% and an F1-score of 95%. Using Shapley Value, we extract key variables that positively of negatively impact the dependent variable(good/defective product). We identify variables with relatively low importance, suggesting variables that should be prioritized for management. Conclusion: The extraction of key variables using SHAP provides explanatory power distinct from traditional machine learning techniques. This study holds significance in identifying key variables that should be prioritized for management in CNC machining process. It is expected to contribute to enhancing the production quality of the CNC machining process.

Because of the rapidly changing environment and high uncertainties, the semiconductor industry is in need of appropriate forecasting technology. In particular, both the cost and time in the test process are increasing because the process becomes complicated and there are more factors to consider. In this paper, we propose a prediction model that predicts a final "good" or "bad" on the basis of preconditioning test data generated in the semiconductor test process. The proposed prediction model solves the classification and regression problems that are often dealt with in the semiconductor process and constructs a reliable prediction model. We also implemented a prediction model through various machine learning algorithms. We compared the performance of the prediction models constructed through each algorithm. Actual data of the semiconductor test process was used for accurate prediction model construction and effective test verification.

본 연구는 실제 원심주조 생산 라인에서 수집된 공정 데이터를 활용하여 수축 기공 불량 예측을 위한 머신러닝 기반 모델을 설계하고, 예측 결과를 바탕으로 주요 공정 인자를 분석하였다. 원심주조 공정에서는 주입 조건, 회전 조건, 온도 및 시간 변수들이 복합적으로 작용하여 단일 변수 기반의 불량 판별에는 한계가 존재한다. 따라서 공정 변수 간의 상호작용을 효과적으로 학습할 수 있는 CatBoost와 XGBoost 기반의 트리 모델을 적용하였다. 실제 생산 데이터는 불량 비율이 낮은 클래스 불균형 문제가 존재하며, 이에 따라 불량 미검출 (FN)과 정상 오검출 (FP)의 비용 차이를 반영한 비용 민감 임계값 (Cost-sensitive Threshold) 최적화를 수행하였다. 또한 단일 모델의 예측 편향을 완화하고 성능을 개선하기 위해 두 모델의 예측 확률을 가중 결합한 앙상블 전략을 적용하였다. 실험 결과, 앙상블 모델은 단일 모델 대비 FN과 FP 간의 균형 측면에서 향상된 성능을 보였다. 또한 변수 중요도 분석 결과, 상위 중요 변수의 대부분이 열적 상태와 관련된 변수로 구성되어 있었으며, 공정 조건 변수 역시 불량 예측에 유의미한 영향을 미치는 것으로 확인되었다. 본 연구는 실제 공정 데이터 기반의 불량 예측과 공정 인자 해석을 통합한 데이터 기반 공정 관리 프레임워크를 제시한다.

In this study, we developed a machine learning-based model for predicting shrinkage porosity defects using process data collected from an actual centrifugal casting production line. An analysis was then performed on key process factors based on the prediction results. Centrifugal casting is a complex process in which pouring conditions, rotational conditions, temperature, and timerelated variables interact simultaneously, making it difficult to reliably discriminate defects using single-variable threshold rules. Therefore, tree-based models, CatBoost and XGBoost, were applied to effectively learn interactions among process variables. Since the defect rate in real production data is low, a class imbalance problem exists. Accordingly, a cost-sensitive decision threshold reflecting the cost difference between false negatives (FN) and false positives (FP) was optimized. To mitigate prediction bias inherent in single models and improve performance, a weighted ensemble strategy combining the predicted probabilities of the two models was applied. Our experimental results show that the ensemble model achieves improved performance in balancing defect detection and false alarms compared with individual models. A feature importance analysis indicates that most of the top-ranked variables correspond to thermal state-related variables, while process condition variables also have a significant influence on defect prediction. This study proposes a data-driven process management framework that integrates defect prediction and process factor interpretation based on real industrial process data.