Alp Akcay

1 Moving a new manufacturing technology from development to a scaled-up manufacturing system brings companies technical, financial, and organizational challenges. Within established production indu...

Business processes are vital to an organisation. Occasionally, those processes lead to undesired outcomes such as delays, quality issues, excessive costs, or low throughput, often caused by complex dependencies between process conditions and resources. Identifying the underlying causes of the process outcomes enables targeted process improvements, a form of post-hoc decision support in which, given an unfavourable completed execution, the system proposes a small set of high-impact, evidence-based changes to improve future process outcomes. Counterfactual explanation is a promising technique to provide such decision support. However, existing techniques lack the ability to deal with constraints specific to processes with complex dependencies. In this work, we propose a counterfactual decision support approach leveraging the object-centric process modelling paradigm to explain how a process execution must be modified to change its outcome to a desired one. The resulting explanations are constrained by the underlying process and domain rules to ensure that the suggested explanations are feasible, actionable, and minimally invasive. The proposed approach is evaluated on multiple synthetic and real-world process datasets and benchmarked against a method for generating generic graph counterfactuals. The results demonstrate that our approach is effective in generating minimal counterfactual explanations and outperforms the tested competitor.

With the advent of digital transformation, organisations are increasingly generatinglarge volumes of data through the execution of various processesacross disparate systems. By integrating and aggregating data from theseheterogeneous sources, it becomes possible to derive new insights essential fortasks such as monitoring and analysing process performance. Typically, thisinformation is extracted during a data pre-processing or engineering phase.However, this step is often performed in an ad-hoc manner and is time consumingand labour-intensive. To streamline this process, we introducea reference model and a collection of patterns designed to enrich productionevent data. The reference model provides a standard way to store andextract production event data. The patterns describe common data aggregationand information extraction tasks. The reference model is constructed bycombining the ISA-95 industry standard with the Event Knowledge Graphformalism. The patterns are developed based on empirical observations fromevent data sets originating in manufacturing processes and are formalisedusing the reference model. The ability of the proposed patterns to acceleratedata aggregation and facilitate knowledge exchange has been validatedthrough an empirical evaluation involving manufacturing professionals fromdifferent roles and production environments.

Accurate endpoint detection is critical in semiconductor plasma etching. Optical emission spectroscopy data contains information about the endpoint, but its analysis requires significant domain knowledge, and it contains large noise when the open area ratio is small. This paper proposes an interpretable data-driven endpoint detection method for small open area etching, leveraging transfer learning. First, we propose a formal metric quantifying the significance of endpoint trend in each wavelength signal and select the wavelength signals that are most sensitive to endpoint variations. Second, we devise an asymmetric autoencoder to uncover the endpoint trend in the noisy signals. Its asymmetric architecture incorporates nonlinear characteristics while ensuring the latent feature reflects the endpoint trends of the individual signals. Experimental results show that the model that learned the large open area etching (11.1%) can detect the endpoints of the small open area etching (0.5%) with the relative mean absolute error less than 0.6%, and amplify the signal-to-noise ratio by a factor of 2-3. Furthermore, the analysis of the selected wavelengths provides deeper insights into the underlying physical processes. The proposed method can be applied with minimal domain knowledge, while its results allow for exploring physical interpretations.

The implementation of Industry 4.0 technologies, especially big data infrastructures, enables easy access to large amounts of data from all stages of semiconductor manufacturing. In this study, we propose a novel methodology to reduce a large amount of sensor data collected from wire bonding machines, called machine signals, into a set of interpretable features relevant to the wire bond quality of chips. The methodology is applied to real-world operational machine signal data from the wire bonding process at NXP Semiconductors N.V. During the wire bonding process, several sensors actively monitor the process, generating signals as discrete multivariate time-series for each semiconductor device. The proposed methodology consists of the following steps. First, we extract features from the discrete multivariate time-series and train a baseline model with all features. Second, we use permutation feature importance to rank the relevance of signals and corresponding features with the goal of identifying the optimal signal and feature set. Finally, the classification performance when using the optimal signal and feature set is compared to the performance of the baseline model. We conclude that the dimensionality of the data can be significantly reduced without losing classification performance. The reduced dimensionality leads to highly interpretable classification results in a real-world wire bonding use case.

In biopharmaceutical manufacturing, fermentation processes play a critical role in productivity and profit. A fermentation process uses living cells with complex biological mechanisms, leading to high variability in the process outputs, namely, the protein and impurity levels. By building on the biological mechanisms of protein and impurity growth, we introduce a stochastic model to characterize the accumulation of the protein and impurity levels in the fermentation process. However, a common challenge in the industry is the availability of only a very limited amount of data, especially in the development and early stages of production. This adds an additional layer of uncertainty, referred to as model risk, due to the difficulty of estimating the model parameters with limited data. In this article, we study the harvesting decision for a fermentation process (i.e., when to stop the fermentation and collect the production reward) under model risk. We adopt a Bayesian approach to update the unknown parameters of the growth-rate distributions, and use the resulting posterior distributions to characterize the impact of model risk on fermentation output variability. The harvesting problem is formulated as a Markov decision process model with knowledge states that summarize the posterior distributions and hence incorporate the model risk in decision-making. Our case studies at MSD Animal Health demonstrate that the proposed model and solution approach improve the harvesting decisions in real life by achieving substantially higher average output from a fermentation batch along with lower batch-to-batch variability.

Brainport Industries Campus (BIC) is a unique initiative taken by the high-tech suppliers of the Eindhoven region to come together and work towards collaborative consumption of shared resources such as production machines, automated guided vehicles, clean rooms and storage spaces in order to jointly meet their customers’ demands, reduce cost of ownership and better manage risks during uncertain demand and supply cycles. However, the matching-demand-with-supply process in a sharing economy requires novel perspectives and tools. In this chapter, we propose a digital platform for serving multiple tenants of BIC sharing a fleet of heterogeneous automated guided vehicles (AGVs). The proposed platform includes a user interface as an order input system, an order database and a cloud server that houses the necessary intelligence for communicating with all the aforementioned modules. This platform that serves multiple tenants sharing a heterogeneous AGV fleet is the first of its kind.