Skip to content

Machine Learning for Failure Analysis

Template for new Applications

A framework has been developed to ensure reproducible, non duplicate, and clean code for projects using machine learning for failure analysis.

ML framework overview

The framework contains domain specific and domain independent building blocks. While domain specific blocks need to be adjusted for each project separately, the domain independent blocks can be reused for all projects. The domain independent blocks, consists of abstract building blocks and can be extended if additional functionality is required. Find enclose a description of the existing building blocks.

  1. Domain specific: Transformation

    • Converter: transform source data into a more handy data format (eg. HDF5)
    • Gatherer: gather hdf files and creates external links for fast queries
  2. Domain independent: Modelling

    • Handler: extracts features using tsfresh, creating context data file with features and meta data to select dataset efficiently
    • Dataset Creator: loads dataset for machine learning
    • Augmentor: augment data to improve results of classifier
    • Classifier: different models for supervised machine learning
    • Explainer: explain results of classifier
    • Anomaly Detector: different models for unsupervised machine learning

Existing machine learning projects are described bellow.

CLIC RF breakdown prediction

Radio Frequency (RF) breakdowns are one of the most prevalent limits in RF cavities for particle accelerators. During a breakdown, field enhancement associated with small deformations on the cavity surface results in electrical arcs. Such arcs degrade a passing beam and if they occur frequently, they can cause irreparable damage to the RF cavity surface. We propose a machine learning approach to predict the occurrence of breakdowns in CERN’s Compact LInear Collider (CLIC) accelerating structures. We use state-of-the-art algorithms for data exploration with unsupervised machine learning, breakdown prediction with supervised machine learning, and result validation with Explainable-Artificial Intelligence (Explainable AI). By interpreting the model parameters of various approaches, we go further in addressing opportunities to elucidate the physics of a breakdown and improve accelerator reliability and operation.

The figure bellow shows the results of a unsupervised machine learning method called 2d-tSNE, used to summarize the available data of this project [1]. Each point indicates an event during which data from many different sensors was stored. The algorithm is able to summarize the data in all events which enables to distinguish between (a) different operational runs r, (b) not filtered and filtered events, and © breakdown and healthy events. In (d) no clear separation of events with a upcoming breakdown and a healthy event in the next pulse is possible. All representations in © are a subset of not filtered events in (b) and all representations in (d) are a subset of all healthy signals in ©.

CLIC RF breakdown prediction

References