Anomaly detection in data sets generated by the CERN Radiation Monitoring Electronic system CROME to develop predictive maintenance algorithms

Several experiments and particle accelerators at CERN, the largest laboratory of high-energy particle physics in the world, require accurate monitoring of ionizing stray radiation to protect people and the environment. To meet this requirement, the new in-house developed and produced radiation monitoring system CROME is responsible for reliable monitoring of the ambient dose rate and permanent data logging, among others. Data-driven methods enable planning of maintenance actions based on failure predictions to keep failure rates of safety-critical systems such as CROME at an extremely low level and ensure high system availability and reliability. The goal of this thesis is therefore to evaluate the usefulness of the datasets generated by CROME devices for failure prediction and to examine the applicability of selected anomaly detection algorithms. More precisely, an anomaly detection process to identify failure precursors and malfunctions is presented. Detected anomalies can thus be used for condition monitoring or system improvement. The proposed process combines noise extraction using wavelet transform and unsupervised anomaly detection algorithms to increase the detectability of the anomalous system behavior. Moreover, statistical features describing the noise and other signal characteristics were used as input to the employed isolation forest algorithm. This model-based algorithm detected samples with unusual signal shapes such as spikes or zero crossings of the dose rate signal. In addition, a long short-term memory autoencoder model was employed to learn temporal dependencies in the measurement observations. Samples considered as anomalous by the autoencoder showed unexpected correlations or unusual measurement values. Nevertheless, system experts still need to evaluate whether the anomalies detected are actual failure precursors or solely infrequent occurrences to make the results useful for the development of predictive maintenance strategies. Due to missing guidelines in the literature on how to configure the wavelet transform, a signal classification process was developed to compare various configurations by analyzing the performance of a supervised machine learning algorithm. This algorithm was trained to detect synthetically modified samples based on statistical measures calculated for extracted noise signals. The proposed signal classification process allows the selection of the most appropriate configuration of the wavelet transform for a given noise extraction use case by relying on real signal samples and providing possibilities to customize the signal modification. Additionally, two approaches to integrate new data in the anomaly detection process, either before or after the initial training of the model, were presented. Based on these findings, a feedback loop allowing the integration of system knowledge was proposed for further development of the presented anomaly detection process.