Speaker
Description
Pulse timing is an important task for nuclear radiation detectors and widely applied in nuclear spectroscopy, radiation imaging, high-energy physics, etc. While neural networks emerge as high-performance alternatives for precision timing of detector signals, the requirement of abundant labelled data poses a challenge for traditional supervised learning and limits the application of such methods. To alleviate the algorithmic thirst for labelled data, in this abstract we combine intra-sample self-supervised learning and outer-sample weakly-supervised learning to form a new optimization paradigm for single-channel pulse timing. For the self-supervised part, we use the intrinsic timing correlation between pairs of waveform segments, along with a random shift for regularization, to learn linear models on arrival time; for the weakly-supervised part, we use extrinsic timing labels on a few waveform subsamples to form reference examples, rectifying mismatches between timing bases of different examples caused by intra-sample supervision. Preliminary results show that, with as few as 1 reference example per 16 self-supervised pairs, the neural network model successfully recovers the pulse arrival time from the sampling points of the waveform digitizer, with a clearly unified time base. We conduct simulation with TensorFlow (Keras) deep learning framework for 3 different waveform variations (fixed, discrete, and continuous), with a 1 GS/s digitizer under signal-to-noise ratio of 54.8 dB. The proposed method overcomes the floating base issue when waveform parameters vary, and achieves timing resolution of 0.05 ns (for the fixed), 0.57 ns (for the discrete), and 0.12 ns (for the continuous), respectively.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | No |