Session: 14-03-01: Reliability and Safety in Industrial Automation Systems

Paper Number: 69390

Start Time: Monday, 05:30 PM

69390 - An Overview of the Research Landscape in the Field of Safe Machine Learning 

Since Machine Learning techniques (ML) have the ability to deal with complex, large and heterogeneous datasets, their application in various fields of industrial automation and across all levels of the automation pyramid can generate a great benefit concerning e.g. efficiency and productivity. First applications can already be found in the field of process optimization, condition monitoring, predictive maintenance, quality control and failure detection. However, with an integration of ML components in industrial automation environments, the overall system safety can be effected. This is obvious, when ML techniques are used as safety functions, but can also be present, when ML techniques are deployed as normal operation functions. In both cases, the impact of ML components has to be assessed and understood in an analogue manner as for any other software and control function with critical influence on safety and according to today’s standards and regulations. Challenging in this regard is that in comparison with classical software and control functions, ML techniques are characterized by high complexity and a dynamic and probabilistic nature.

In Europe the Machinery Directive 2006/42/EC defines requirements for machinery and certain parts of machinery with the intent to ensure a defined safety level. One of these requirements is, that “the manufacturer or his authorised representative should also ensure that a risk assessment is carried out for the machinery which he wishes to place on the market”. An iterative approach for assessing these risks is given in the international standard ISO EN 12100. A part of the risk assessment is the risk evaluation, which includes risk estimation. To be able to carry out such a risk estimation, a thorough understanding of each component of the machinery, including the software part, and their interplay is mandatory. This implies the precise description of e.g. the reliability of the respective ML components.

In principle, there exist 2 basic routes to demonstrate the reached safety level of a complex system including software components:

·     -   management of the realization lifecycle, where techniques and measures are defined or

·     -   technical assessment, where aspects like dependability or risks of a complex system are evaluated, a model is generated and analyzed.

This article aims at giving an overview of existing approaches for the safety assessment of ML algorithms by categorizing them, visualizing the scientific landscape in this field and discussing white spots. This is done by a systematic literature review in combination with text mining and a quantitative text analysis.

Presenting Author: Georg Siedel Federal Institute for Occupational Safety and Health

Authors:

Georg Siedel Federal Institute for Occupational Safety and Health
Stefan Voß Federal Institute for Occupational Safety and Health
Silvia Vock Federal Isntitute for Occupational Safety and Health