Anomaly detection using an Autoencoder for the High-Granularity Calorimeter

The High-Granularity Calorimeter (HGCAL) is the new upcoming endcap for the Compact Muon Solenoid (CMS) detector, replacing the existing endcap calorimeter for the High-Luminosity Large Hadron Collider (HL-LHC) era. The HGCAL contains silicon sensors as active elements, which performs the task of signal detection. This brief Summer Student project seeks the investigation of Machine Learning based anomaly detection tools for data quality monitoring of the HGCAL. The implementation takes place in the PyTorch environment, using processed pedestal data to train an Autoencoder. By using the CERN Service for Web based Analysis platform "SWAN", efficient hardware acceleration using CUDA is possible. Using an unsupervised deep learning approach, an autoencoder-based anomaly detection system has been developed with predefined clean background data. The AE is able to perform the reconstruction between the input- and output layer from the prepared training data. This summer student project provides a first approach for an anomaly detection development.