Real-Time Wiener Deconvolution for feature reconstruction in JUNO

Abstract

In particle physics, experiments generate substantial amounts of data that can be difficult to process without preliminary scaling. To avoid losing potentially crucial data, experimental collaborations are studying novel techniques for real-time data processing to extract features for further physics analysis. A common approach, especially in neutrino physics, is to use FPGAs for data acquisition and pre-processing. This paper presents an advanced Real-Time Wiener deconvolution algorithm designed to leverage the processing capabilities of the FPGA integrated into the readout boards of the Jiangmen Underground Neutrino Observatory (JUNO). The goal is to enable real-time reconstruction of the signal generated by photomultiplier tubes (PMTs) when neutrino interactions are detected. By exploiting online reconstruction of the signal generated by PMTs, we expect to improve the detection of low-energy depositions, such as those produced by transient astrophysical phenomena. These depositions are usually not saved because of the significant background that affects the low end of the energy spectrum, which would result in a large trigger rate, hence a large amount of data required for storage. This paper presents the features of the algorithm, including its ability to manage high-throughput data streams with minimal latency, adaptability, and resilience in discerning the characteristics of input data. Performance is evaluated on a JUNO electronic board. This study further demonstrates the potential of FPGA-based solutions for neutrino physics.

Figures (18)

• Figure 1: Experimental setup. The experimental setup is composed by a CAEN detector emulator, able to simulate arbitrary output signal, a connected to the emulator via a coaxial cable and an oscilloscope connected in parallel to monitor the signal shape.
• Figure 2: Data acquisition chain. The data acquisition chain can be schematized as follows: a detects light and accumulates charge (green); the analog output signal (orange) is digitized by a Flash and processed by the ; the digitized signal (blue) and the reconstructed features (time and charge) of the hits are sent to the . The vertical range of the last two panels is normalized to ease the comparison between the analog and the digital signals.
• Figure 3: Template construction steps. (a) After removal of baseline, single photoelectron waveforms are selected to represent the impulse response of the . (b) The rising edges of the peaks are aligned, and the average amplitude is computed for each time sample. (c) The resulting average waveform constitutes the template of the response for our setup.
• Figure 4: Power Spectral Density characterization. (a) The Power Spectral Density of the signal template shown in \ref{['fig:template']}. The peak at 250 MHz is a distortion caused by the detector emulator during signal generation. (b) The noise Power Spectral Density is characterized by white noise—constant value across the frequency domain—and specific colored noise spikes.
• Figure 5: filter frequency response comparison. The filter response deviates from the ideal frequency response for both the Wiener (a) and Deconvolution (b) filters. This discrepancy results from the approximation inherent in the method when minimizing error with a limited coefficient count. The dips observed in the FIR response in panel (a) correspond to the zeros of the filter's frequency response. However, this is not an issue, as the objective is to construct filters that exhibit the desired behavior in the time domain.