Speaker
Description
The complexity of the ATLAS detector and its infrastructure require an excellent understanding of the interdependencies between the components when identifying points of failure (POF). The ATLAS Technical Coordination Expert System features a graph-based inference engine that identifies a POF provided a list of faulty elements or Detector Safety System (DSS) alarms. However, the current algorithm can be very time-intensive and often returns an impractically large list of circa 100 potential POFs to the operators. Here, we present a new tool that, based on DSS alarms, instantly predicts the class of the potential single POF and allows the assessment of the criticality of the failure. Currently, the tool predicts 28 classes of single POFs, including Flammable Gas, Smoke detection, and cooling systems. It comprises several one-versus-rest decision-tree classifiers, each instantly estimating the correspondence of the POF to a class. The classifiers are trained on a simulated dataset obtained with the Expert System, and provide a fully understandable white box model that is required for safety-critical applications. Their performance, foremost the false-negative rate, is evaluated by leave-one-out validation. For 8 (16) classes, we achieve a false-negative rate of less than 1% (5%) and a true-positive rate of more than 96%. This corresponds to an accuracy of more than 85% for the five best performing classes. Additionally, the training data is used as a look-up table that maps alarm states to a single POF. The simulation shows that about 60% (85%) of the simulated alarm states are triggered by one (two) unique element(s).
| Consider for long presentation | No |
|---|
