Speaker
Description
In a multi-channel radiation detector readout system, waveform sampling, digitization and transmitting bits to the data acquisition system constitutes a conventional processing chain. The quantities, such as time-of-arrival and signal magnitude, i.e. deposited energy is estimated by fitting analytical models over the acquired digital data extracting starting times of signals, peak amplitudes, or areas under pulse envelopes. However, such a data processing could be carried out through machine learning algorithms on the front-end ASICs often termed as edge computing. Edge computation offers enormous benefits, especially when the analytical forms are not fully known, or the registered waveforms suffer from imperfections of practical implementations. Also, experimental findings early in the data processing chain often reduces bandwidth of the data throughput, thereby reducing overall cables in any of the experiment.
Our study has focused on investigation of various neural networks and their implementation, training them with sensor signals of single peak from a single-channel typical from a silicon sensor for predicting peak amplitude. The sampling rate has been reduced, resulting 3 to 4 sample points on the signal peak. We have investigated two types of neural network: MultiLayer Perceptron (MLP) and 1D-Convolutional Neural Network (CNN) till date. Both are popular choices for common machine learning tasks and offer different characteristics of computation complexity and parameter efficiency. These neural networks are optimized in terms of hidden layers, neurons and weights through pruning while still maintaining an acceptable inference accuracy. In this workshop, we would like to present the overall performance and hardware requirements for the Multi-Layer perceptron (MLP) and the Convolutional Neural Networks (CNN) in predicting the peak amplitude of a single channel sensor response.
