Pulse Recognition by using a Convolutional Neural Network (CNN)

The strategy aimed at the noise reduction is a classical topic in electronics, and its reduction is usually achieved in the frequency domain, making appropriate Fourier and Wavelet transform in combination with appropriate filtering. Their main disadvantages may be represented by the possible distortions of the physical contents contained in the filtered signals that in several condition need to be completely preserved. We developed an innovative technique based on Convolutional Neural Networks (CNN) to manage the charge pulses produced by the particle silicon detectors. This algorithm is capable to identify pulses produced by true physical events with an accuracy of more than 99%, even in presence of a significant amount of noise with no filtering process. The training phase of the algorithm requires a selected set of pulses individually labeled as true, in order to classify the data and thus predict outcomes in a classic supervised learning algorithm. The combination of different techniques and algorithms improve the quality of the result and can represent a new and interesting frontier for the future of data analysis.