Algorithmic Hit Filter:
Removes easily recognizable noise patterns, such as Delta-rays, cell-edge hits, and electronic noise. They can be characterized as such:
- Delta rays: The signature of a delta ray moving along the X-Y plane is a cluster consisting of more than two continuous hits with large average TDC-time differences (10 ns or larger). The two hit edges are stored and the middle hits are discarded in order to maintain viable hits that could have been generated by a muon track.
- ”Cell-Edge” hits: A track passing near the center of two adjacent wires may induce ionizations that fire both wires. This results in the respective hits having long drift distances (about half of the cell width). The hit with the longest drift distance is therefore removed.
- Electronic Noise: If two or more hits on neighboring wires have average TDC time differences less than 10 ns, they are discarded as electronic noise.
These can be easily removed using if statements.
Event Filter:
After the algorithmic hit removal, we train a classification neural network to recognize hit patterns of interesting physics events. The network evaluates an event and outputs a probability that the event is worth further analysis.
Some characteristics the network can be trained to look for are:
- Containing any reconstructable muons
- Containing two reconstructable muons
- Containing a reconstructable dimuon pair
- Containing a reconstructable dimuon pair from the target
Depending on the desired characteristic and the probability threshold we assign, more or fewer events will be let through for further analysis.
Classification neural networks often have three measures of performance: precision, recall, and F-measure. The precision is defined as true positives divided by assigned positives. The recall is defined as true positives divided by true positives plus false negatives. The F-measure is the harmonic mean of the two.
The event filters trained so far have an F-measure of around 0.8.
Track Finder
Events deemed worthy of further analysis are passed to a neural network that determines which hits correspond to full dimuon tracks. This network uses convolution layers to