How to exploit system data for operational excellence?

Emerging operational data, a direct consequence of systems increasingly being equipped with sensors and software, could help industry to deal with the key challenges in industry’s shift towards complex digitised, connected, intelligent solutions. How to make these systems resilient to change, capable to autonomously self-adapt their operations upon unforeseen conditions? Will augmenting systems with introspective and machine-learning capabilities be the next step in supporting evolvability? What is needed to make self-learning systems genuine artificial intelligent? These and many other questions are being discussed in this session, with a focus on pursuing operational excellence by means of self-learning behavior.

Moderation: Bas Huijbrechts, ESI (TNO)

Chair: Gregor Pavlin, Thales TRT

The recent proliferation of increasingly advanced sensors and software in complex industrial systems has enabled access to large amounts of data that can provide key insights about industrial processes. Automated interpretation of such data and system control based on non-trivial models and algorithms can make industrial processes more efficient, cleaner and safer. Powerful machine learning (ML) techniques in principle enable (i) automated extraction of sophisticated models from the data and (ii) continuous adaptation of such models to changing operational conditions. While impressive results have been obtained in many areas, the use of ML in complex mission critical industrial applications remains very challenging. Firstly, due to the nature of ML, it can be difficult to guarantee the learning processes produce adequate models. Secondly, the validation of increasingly complex ML-based solutions can be labour intensive, time consuming and even intractable if approached in a naïve way. The key to efficiently addressing these challenges are enhanced development processes and training for engineers that adequately consider the underlying ML principles.

download pdf

Michael Borth

Here be Dragons; Charting the Engineering of Learning Systems

We move towards the realization of smart systems. Information and its underlying data forms the backbone for many such cyber-physical systems, opening the possibility of system’s that observe and learn and thus adapt. But as they improve their fit to their operational context, current tasks, or their user’s preferences, they change.

In this talk, I test our understanding of the scale this change may take, investigating if we, the professionals in engineering, are prepared to accept that the system which operates is no longer the one we designed. Do systems that truly learn bring us into uncharted fields? Are there dragons in the white spaces of the engineer’s maps if, e.g., tests and validation can only lack behind and past observations mean little for diagnosis or prognosis of novel situations? The lack of control and complexity that is induced by adaptivity more than hints at a yes to these questions – and the need to up our science with probabilistic predictions, architectures for reflection and resilience, and risk-aware optimizations

download pdf

John van den Dobbelsteen, TU Delft

Computer assisted process management in the operating room

The trend towards personalized treatment results in an increasingly complex technical environment in the clinical pathway, leading to a growing workload for health care professionals. Algorithms that forecast and mitigate clinical risks based on complex situation analyses may enable partial automation and management of tasks. However, the lack of methods to incorporate medical knowledge and data into models hinder the development and implementation of intelligent assistance technologies.

To enable robust data acquisition under specific constraints in the operating room, major hurdles must be overcome. These are related to both the integration of novel sensor technologies with the therapeutic instruments and the processing of the data into meaningful descriptions of clinical events. Yet, existing medical systems have very limited level of integration, are monolithic and proprietary. Due to its proprietary character, interdisciplinary-integration is not a standard component of current medical engineering or computer science training. To address this gap, machine learning-interpretable representations of domain knowledge concerning OR documentation, resource management, and perioperative business processes need to be developed and validated for streamlining care path planning.

download pdf

go to S3 Architecting intelligent systems