In the industry there is a need for guaranteed quality under all conditions. Both in mechatronic products, in production processes as in assembly lines. Unfortunately, failures happen and pre-established requirements are not met as a consequence of:
- Yet undiscovered intra cycle conditions (e.g. environmental conditions);
- A sequence of inter cycle conditions (e.g. accumulation of errors);
- Time degradation of the system over cycles.
The technological barriers we encounter are:
- Detect: The difficulty of obtaining the relevant data of scarce events with numerous and various inputs/parameters;
- Predict: The absence of a framework to understand and quantify how conditions lead up to a quality degradation (e.g. failure event);
- Mitigate: The absence of an explainable decision support tool to define and evaluate on-and offline mitigation strategies.
Bayesian networks (BN) provide a strong framework, i.e. a probabilistic graphical model to represent conditional dependencies (causalities) between root causes and quality (risk) events. BN allow to propagate probabilities of root causes to risk/quality events; and back-propagate (infer) occurring risks to root causes to uncover the relationships. To make these applicable for industry, the project will tackle the following sub goals:
- The graphical model describes dependencies between parameters, states, … but how such a graphical model can be built or leverage on prior knowledge (e.g. physical models, heuristic relations) is not yet established.
- Training data to quantify the relations in the graph is limited as data is “big-scarce”: there is a lot of data on the nominal operation; there are numerous and various inputs/parameters; but there is few data on the failing events.
- In mechatronic systems drift of parameters due to degradation are often a key underlying reason for quality issues. However, Bayesian networks cannot yet cope well with such ‘concept drifts’.
- Not all root causes can be explained by static parameter values/features; also quality events are embedded in the dynamic profile in time. BN hence, need to be able to deal with dynamic conditions/behavior.
- Method to effectively exploit such a Bayesian network to turn it into the system intelligence in order to detect, predict and mitigate; this requires systematic methods for approximating distributions, defining cost models, sampling, ...
QUASMO is a Strategic Basic Research (SBO) project. We are looking for companies to join the User Group and work with us on the valorisation of the project.
Interested? Complete the form below and we will contact you as soon as possible.