Neural Online Filtering Based on Preprocessed Calorimeter Data

Among LHC detectors, ATLAS aims at coping with such high event rate by designing a three-level online triggering system. The first level trigger output will be ~75 kHz. This level will mark the regions where relevant events were found. The second level will validate LVL1 decision by looking only at the approved data using full granularity. At the level two output, the event rate will be reduced to ~2 kHz. Finally, the third level will look at full event information and a rate of ~200 Hz events is expected to be approved, and stored in persistent media for further offline analysis. Many interesting events decay into electrons, which have to be identified from the huge background noise (jets). This work proposes a high-efficient LVL2 electron / jet discrimination system based on neural networks fed from preprocessed calorimeter information. The feature extraction part of the proposed system performs a ring structure of data description. A set of concentric rings centered at the highest energy cell is generated for each calorimeter layer, producing a total of 100 ring. Energy normalization is later applied to the rings, making the proposed system usable for a broad energy spectrum. For data compaction and signal decorrelation, a segmented preprocessing scheme is developed. For this, Principal Component Analysis (PCA) and Principal Component of Discrimination (PCD) are compared in terms of compaction rates and classification effici ency. Moreover, as signal decorrelation is performed either linearly (PCA) or nonlinearly (PCD) in the preprocessing phase, the classifier complexity can be reduced, which also favors signal preprocessing adoption in level two trigger strategy. For the hypothesis testing section, an artificial neural network was employed. Neural networks are fast to execute, and can provide non-linear cuts in a high dimensional space, providing better pattern separation for complex problems when compared to linear decision systems.