Previous research clearly showed the potential of video analysis and machine learning to study dairy cow behaviour. However, no long term study with the aim of collecting and analysing behavioural data of a commercial sized herd was yet executed. Therefore, we aim to develop a video analysis based monitoring system to study the behaviour dairy cows in an automated way. With this computer vision based pipeline, we will carry out a profound study of the behaviour of dairy cows in commercial herds, both in indoor and grazing conditions.

The study of animal behavior allow scientiests to percieve the individual and social patterns. Although it is a rapidly growing area of study, the traditional approaches still predominate from collection of behavioral data to analysis of it. By creating automated pipelines and developing validation strategies for species under the focus, we aim to develop methodologies for characterisation of caged-animal behaviors. Application of state-of-art computer vision algorithms and integrating prior knowledge about the environmental knowledge, we build accurate and robust techniques from collection of data to knowledge extraction from it.

The development of advanced analytics for monitoring and decision support is essential in the context of automation and (sensor) technology. Oftentimes, these analytics serve to extract information from time series data, in which certain measurements are taken from a target subject. In agriculture, these target processes are complex and often influenced by a multitude of internal (e.g. genetics, physiology) and external (e.g. contexts, management) factors. I hypothesize that the integration of expert knowledge in the analytics renders the interpretation more precise, accurate, timely and useful for decision support, even when different states of the target are not discrete or their severity (and thus, the required action) depends on a combination of different factors. In this project, I aim to develop a general framework for the integration of 3 types of expert knowledge into state-estimation time series analytics, and apply them on 2 agriculture-related cases. I will focus on the translation of domain expertise, historical data patterns and context measures to incorporate them in time series models targeting monitoring heat stress in dairy cows and welfare in intensive fish production systems. The general framework will provide a basis for how to consider different types of expert knowledge in biological and agricultural production systems, and via the case studies, provide leverage for more sustainable primary production.

During development of new drug products, both the formulation and production process need to be optimized to achieve robust performance and deliver a final drug product that meets a series of critical quality attributes (CQAs). Formulation development is in many cases based on the formulator’s experience by means of trial-and-error which can potentially lead to increased experimental workload which induces additional costs and delayed time-to-market. This project aims to implement machine learning techniques into the drug development process to guide the experimentation during drug development which will can lead to a reduction of the number of experiments that is needed to obtain a successful final drug product.

Advanced Therapy Medicinal Products (ATMPs) will revolutionize 21st century medicine, as evidenced by several major scientific and clinical breakthroughs in the past decade. By using cells as therapeutic agents, one can treat or cure an unprecedented variety of diseases or defects, even as a personalized therapy. However, the complexity of living cell products makes the ATMP field equally challenging as it is promising. Of the 10 ATMPs approved by EMA from 2009-2018, only 6 are currently still on the market, which is a waste of public resources as well as potentially life-changing therapies. At present, ATMPs are typically not effective enough, as well as too expensive. Personalized, closed and automated processes (i.e. industry 4.0) are often cited as a key facilitators towards the mitigation of these challenges. A prerequisite for these processes, as corroborated in the Process Analytical Technology (PAT) initiative and statement by the FDA, is that the Critical Quality Attributes of the product can be evaluated cost-effectively, non-invasively, automatically, and in real-time. However, there are no off-the-shelf sensor options for ATMPs that meet these criteria. More advanced analytical methods such as Raman spectroscopy, impedance spectroscopy and soft sensing have been cited as promising monitoring methods for cell-based processing for over a decade. There are two main challenges to the use of these methods. In this project, we would like to address these challenges. Additionally, we will explore resource-efficient process optimization strategies.

Mathematical modeling of pharmaceutical processes helps scientist to design efficient and high quality production processes. Physics-based models are typically time consuming to perform, therefore hybrid approaches with the data-driven models holds a great potential to increase the use of such models. In our research, we focus on developing surrogate models for the CFD simulations of bioprocesses using machine learning.

Hybrid modeling of surface water systems using online sensor data is a technique used to model the behavior of surface water systems, such as rivers, lakes, and reservoirs, using real-time sensor data. The goal of this modeling technique is to create accurate and reliable models that can be used to make informed decisions about the management of these water systems. In this approach, the modeling process combines the use of physical models and data-driven models. Physical models use mathematical equations to describe the behavior of the water system, while data-driven models use statistical algorithms to identify patterns and relationships in the sensor data. By combining these two approaches, hybrid models can account for both the physical processes that govern the water system and the variations in real-time sensor data.

Hybrid modelling is a solution to bring forth the advantages of both mechanistic and data-driven models. The application of hybrid modelling in the domain of wastewater treatment has the potential to foster automation, increase efficiency, and increase the predictive power of models. The project aims to explore and apply advanced image analysis techniques to describe and predict the complex processes of flocculation and sedimentation in wastewater treatment processes with greater precision. This approach aims to provide a deeper understanding of the fundamental processes while improving the predictive capacity of such systems. This research aims to contribute to the optimisation and efficiency of wastewater treatment technologies.

Our research focuses on extracting structured knowledge from unstructured natural language texts. These texts range from industrial maintenance reports to publications on nature conservation. Various techniques are explored such as knowledge graph construction, failure code prediction, topic clustering and query systems.