Search

In most data analytic applications in manufacturing, understanding the data-driven models plays a crucial role in complementing the engineering knowledge about the production process. Identifying relevant input variables, rather than only predicting the response through some “black-box” model, is of great interest in many applications. There is, therefore, a growing focus on describing the contributions of the input variables to the model in the form of “variable importance”, which is readily available in certain machine learning methods such as random forest (RF). Once a ranking based on the importance measure of the variables is established, the question of how many variables are truly relevant in predicting the output variable rises. In this study, we focus on the Boruta algorithm, which is a wrapper around the RF model. It is a variable selection tool that assesses the variable importance measure for the RF model. It has been previously shown in the literature that the correlation among the input variables, which is often a common occurrence in high dimensional data, distorts and overestimates the importance of variables. The Boruta algorithm is also affected by this resulting in a larger set of input variables deemed important. To overcome this issue, in this study, we propose an extension of the Boruta algorithm for the correlated data by exploiting the conditional importance measure. This extension greatly improves the Boruta algorithm in the case of high correlation among variables and provides a more precise ranking of the variables that significantly contribute to the response. We believe this approach can be used in many industrial applications by providing more transparency and understanding of the process., Validerad;2024;Nivå 2;2024-02-14 (sofila);Full text license: CC BY-NC 4.0

In many industrial applications, obtaining labeled observations is not straightforward as it often requires the intervention of human experts or the use of expensive testing equipment. In these circumstances, active learning can be highly beneficial in suggesting the most informative data points to be used when fitting a model. Reducing the number of observations needed for model development alleviates both the computational burden required for training and the operational expenses related to labeling. Online active learning, in particular, is useful in high-volume production processes where the decision about the acquisition of the label for a data point needs to be taken within an extremely short time frame. However, despite the recent efforts to develop online active learning strategies, the behavior of these methods in the presence of outliers has not been thoroughly examined. In this work, we investigate the performance of online active linear regression in contaminated data streams. Our study shows that the currently available query strategies are prone to sample outliers, whose inclusion in the training set eventually degrades the predictive performance of the models. To address this issue, we propose a solution that bounds the search area of a conditional D-optimal algorithm and uses a robust estimator. Our approach strikes a balance between exploring unseen regions of the input space and protecting against outliers. Through numerical simulations, we show that the proposed method is effective in improving the performance of online active learning in the presence of outliers, thus expanding the potential applications of this powerful tool., Validerad;2024;Nivå 2;2024-02-14 (sofila);Funder: DTU Strategic Alliances Fund;Full text license: CC BY 4.0

Online active learning is a paradigm in machine learning that aims to select the most informative data points to label from a data stream. The problem of minimizing the cost associated with collecting labeled observations has gained a lot of attention in recent years, particularly in real-world applications where data is only available in an unlabeled form. Annotating each observation can be time-consuming and costly, making it difficult to obtain large amounts of labeled data. To overcome this issue, many active learning strategies have been proposed in the last decades, aiming to select the most informative observations for labeling in order to improve the performance of machine learning models. These approaches can be broadly divided into two categories: static pool-based and stream-based active learning. Pool-based active learning involves selecting a subset of observations from a closed pool of unlabeled data, and it has been the focus of many surveys and literature reviews. However, the growing availability of data streams has led to an increase in the number of approaches that focus on online active learning, which involves continuously selecting and labeling observations as they arrive in a stream. This work aims to provide an overview of the most recently proposed approaches for selecting the most informative observations from data streams in real time. We review the various techniques that have been proposed and discuss their strengths and limitations, as well as the challenges and opportunities that exist in this area of research.

Injection molding is one of the most important processes for the mass production of plastic parts. In recent years, many researchers have focused on predicting the occurrence and intensity of defects in injected molded parts, as well as the optimization of process parameters to avoid such defects. One of the most frequent defects of manufactured parts is blush, which usually occurs around the gate location. In this study, to identify the effective parameters on blush formation, eight design parameters with effect probability on the influence of this defect have been investigated. Using a combination of design of experiments (DOE), finite element analysis (FEA), and ANOVA, the most significant parameters have been identified (runner diameter, holding pressure, flow rate, and melt temperature). Furthermore, to provide an efficient predictive model, machine learning methods such as basic artificial neural networks, their combination with genetic algorithms, and particle swarm optimization have been applied and their performance analyzed. It was found that the basic artificial neural network (ANN), with an average accuracy error of 1.3%, provides the closest predictions to the FEA results. Additionally, the process parameters were optimized using ANOVA and a genetic algorithm, which resulted in a significant reduction in the blush defect area.

We present a unified modeling framework for Time-Based Maintenance (TBM) and ConditionBased Maintenance (CBM) for optimization of replacements in multi-component systems. The considered system has a K-out-of-N reliability structure, and components deteriorate according to a multivariate gamma process with L´evy copula dependence. The TBM and CBM models are formulated as Markov Decision Processes (MDPs), and optimal policies are found using dynamic programming. Solving the CBM model requires that the continuous deterioration process is discretized. We therefore investigate the discretization level required for obtaining a near-optimal policy. Our results indicate that a coarser discretization level than previously suggested in the literature is adequate, indicating that dynamic programming is a feasible approach for optimization in multi-component systems. We further demonstrate this through empirical results for the size limit of the MDP models when solved with an optimized implementation of modified policy iteration. The TBM model can generally be solved with more components than the CBM model, since the former has a sparser state transition structure. In the special case of independent component deterioration, transition probabilities can be calculated efficiently at runtime. This reduces the memory requirements substantially. For this case, we also achieved a tenfold speedup when using ten processors in a parallel implementation of algorithm. Altogether, our results show that the computational requirements for systems with independent component deterioration increase at a slower rate than for systems with stochastic dependence.

Implementation of digital, intelligent control systems have traditionally not been easy because of various reasons including harsh temperature conditions, standardized measurements, and conventional mindset in foundries. However, it is possible to digitalize and control the foundry industry by exploiting the modern-day technological tools. The present project presents an integrated physics-based and statistical modelling approach to produce models that forms the fundamental building blocks for digitalization of a continuous casting foundry. Real life online experimentation is a crucial step across the progress of the project to validate the science-based models and to produce data for statistical models.