About the project

Objective
The main objective of this project is to improve the productivity of two packaging lines within the SweOps Steriles function, as measured by Overall Equipment Effectiveness (OEE) and competence in line staffing. We aim to achieve the objective by enabling next-best-action decision support for front-line operators.

Background
Modern pharmaceutical packaging lines are complex systems with multiple intricate physical and digital components. Operators gain domain expertise through extended exposure and interaction with the systems. They see, touch, and listen to the operating parts of the system. With time, they develop deep procedural knowledge and reach the level where they can predict when the physical systems need maintenance. How do they do this? This is the basic question of the project SMART: Smart Predictive Maintenance for the Pharmaceutical Industry, a collaboration between AstraZeneca and KTH.

Our approach deploys three pillars: 1) sensor networks in manufacturing, 2) machine learning predictive models, and 3) interactive immersive and contextual visualizations. We will observe and interview expert operators to acquire their procedural knowledge and focus on the sensing and machine learning tools to produce a rich sensor-based predictive model that we visualize peripherally in the plant and immersively to the operators. The project aims to enhance the operators’ abilities to perform predictive maintenance and expedite the transfer of these skills to novice operators via novel digital tools.

Partner Postdocs

Main supervisors

Background and summary of fellowship
Behaviour Trees (BTs) represent a hierarchical way of combining low-level controllers for different tasks into high-level controllers for more complex tasks. The key advantages of BTs have been shown to include the following:

In this project, we will use the properties of BTs listed above to synthesize controllers that combine the efficiency of reinforcement learning with formal performance guarantees such as safety and convergence to a designated goal area.

Background and summary of fellowship
Reinforcement Learning (RL) is concerned with learning efficient control policies for systems with unknown dynamics and reward functions. RL plays an increasingly important role in a large spectrum of application domains including online platforms (recommender systems and search engines), robotics, and self-driving vehicles. Over the last decade, RL algorithms, combined with modern function approximators such as deep neural networks, have shown unprecedented performance and have been able to solve very complex sequential decision tasks better than humans. Yet, these algorithms are lacking robustness, and are most often extremely data inefficient.

This research project aims at contributing to the theoretical foundations for the design of data-efficient and robust RL algorithms. To this aim, we develop a fundamental two-step process:

  1. We characterize information-theoretical limits for the performance of RL algorithms (in terms of sample complexity, i.e., data efficiency)
  2. We leverage these limits to guide the design of optimal RL algorithms, algorithms approaching the fundamental performance limits

Background and summary of fellowship
In many application areas, it is not sufficient to present the output of machine learning models to the users without providing any information on what leads to the specific predictions or recommendations and how (un)certain they are. The strongest machine learning models are however often essentially black boxes. In order to enable trust in such models, techniques for explaining the predictions in the form of interpretable approximations are currently being investigated. Another cornerstone for enabling trust is that the uncertainty of the output of the machine learning models is properly quantified, e.g., that the output prediction intervals or probability distributions are well-calibrated.

Motivated by collaborations with Karolinska Institutet/University hospital on sepsis prediction, Scania on predictive maintenance and the Swedish National Financial Management Authority on gross domestic product (GDP) forecasting, techniques for quantifying uncertainty and explaining predictions will be developed and evaluated. In addition to scientific papers, the output of the project will be Python packages to support reliable machine learning, enabling predictions of state-of-the-art machine learning models to be complemented with explanations and uncertainty quantification.

Background and summary of fellowship
Social robots and virtual agents are currently being explored and developed for applications in a number of fields such as education, service, retail, health, elderly care, simulation and training and entertainment. For these systems to be accepted and successful, not only in task-based interaction but also to maintain user engagement, in the long run, it is important that they can exhibit varied and meaningful non-verbal behaviours, and also possess the ability to adapt to the interlocutor in different ways. Adaptivity in face-to-face interaction (sometimes called mimicry) has for example been shown to increase liking and affiliation.

This work addresses how style aspects in non-verbal interaction can be controlled, varied and adapted, across several modalities including speech, gesture and facial expression. The project entails novel data collection of verbal and non-verbal behaviours (audio, video, gaze tracking and motion capture) with rich style variation, but also makes use of existing datasets for base training. Synthesis models trained on this data will be based primarily on deep probabilistic generative modelling, conditioned with relevant style-related parameters. Multimodal generation paradigms, that produce congruent behaviours in more than one modality at a time, e.g. both speech and gesture, in a coherent style, will also be explored and evaluated in perceptual studies or experiments with real interactive contexts.

About the project

Objective
The objective of the Swedish Question Generation for Assessing Reading Comprehension (SWE-QUEST) project is to develop a demo system that, given a text, automatically generates multiple-choice reading comprehension questions on the text, as shown in the picture below.

Using a neural-network-based generative model, the demo system will generate the whole multiple-choice question, both the question itself and the answer alternatives. Although the example above was in English, our demo system will work for Swedish text. The project will push the state-of-the-art in natural language generation. The system is intended to be used by teachers of SFI (Svenska för invandrare) to facilitate test construction and the development of teaching materials. Still, it can also be used for self-studies of Swedish.

Background
The SFI students form a heterogeneous group, and the SFI classes are often too big to allow the teacher time to adapt the level of teaching to cater to each student individually. Therefore, developing such a tool could be highly useful. It would rapidly and easily generate several suggested multiple-choice questions (MCQ) on text material for teaching and assessment.

The task of automatically generating reading comprehension questions (without distractors) using neural methods has been studied before, primarily for English. There have also been some attempts at generating distractors, given the question and the correct answer, using neural methods. However, no attempt has been made to construct a trainable model that generates the whole MCQ in one go.

Crossdisciplinary collaboration
The researchers in the team represent the KTH Schools of Electrical Engineering and Computer Science, the Division of Speech, Music and Hearing and the Department of Swedish Language and Multilingualism at Stockholm University.


About the project

Objective
Susan’s ride on Campus2030 aims to demonstrate the potential of digitalization in reducing the carbon footprint and improving the cost-efficiency of the construction and transportation industry. With this objective, the project will establish a one-of-a-kind smart road infrastructure demonstrator on the KTH campus Valhallavägen for the integrated design, construction and operation of smart infrastructures. The demonstrator will incorporate a digital twin of the KTH campus, corresponding to multiple models and data sets that enable virtual assessment and experience of the Campus infrastructure while being validated and updated through real-time data feeds from various sensors. Our work in this direction can be seen on our testbed for Intelligent Transportation Systems on www.adeye.se. Susan’s ride will showcase the potential of edge computing, federated learning, and digital twins in the digital transformation of road construction and autonomous vehicle path planning.

Background
Autonomous vehicles, dynamic charging of electric vehicles and vehicle-to-infrastructure communication are just a few examples that require a systemic solution to function sustainably. Making the smart road sustainable requires a partnership between road owners, operators, electricity companies, vehicle manufacturers, transport and logistics companies, and technology suppliers in digitalization. Data will become a fundamental asset in this partnership. They must be collected through a combination of new sensors in the infrastructure already upon construction on smart vehicles, including construction machinery.

Crossdisciplinary collaboration
The researchers in the team represent the School of Electrical Engineering and Computer Science, KTH, the School of Architecture and the Built Environment, KTH and the School of Industrial Engineering and Management, KTH. The project leverages and extends research carried out in the Campus 2030 project and the TECoSA research centre.