Speaker
Description
This poster presents an advanced real-time Wiener deconvolution algorithm designed to take advantage of the FPGAs integrated into the JUNO experiment readout boards. Exploiting online reconstruction of the signal generated by PMTs, we expect to enable the detection of low energy depositions, like those generated by transient astrophysical phenomena.
The features of the algorithm are presented, including its capacity to manage high-throughput data streams with minimal latency, its adaptability and resilience in discerning the characteristics of the input data. This study further demonstrates the potential of FPGA-based solutions for neutrino physics.
Summary (500 words)
The amount of data generated by particle physics experiments is constantly increasing, with the development of new and improved detectors and acquisition systems. However, processing all this data is not always feasible and the data throughput needs to be reduced through compression or pre-scaling.
The work presented in this poster aims to compress the data coming from the detector acquisitions to save as much information as possible, while using a limited number of resources. To achieve this goal, we developed a Real-Time Wiener Deconvolution (RTWD) algorithm on FPGA; the algorithm is based on Deconvolution, a technique used in signal processing to reconstruct features of a signal f distorted by a generic system h. This technique is usually applied in the frequency domain
F(ω)=(G(ω))/(H(ω)), (1)
through the Fast Fourier Transform (FFT).
Since the quality of the features reconstruction is dependent on the Signal-to-Noise Ratio (SNR) of the distorted signal,
SNR(ω)=S(ω)/N(ω), (2)
Deconvolution is usually paired with the use of a Wiener filter,
W(ω)=SNR(ω)/(SNR(ω)+1), (3)
a filter designed to maximize the SNR of a signal, if the noise is known to some extent.
This algorithm is known as Wiener Deconvolution:
D_W (ω)=1/H(ω)⋅SNR(ω)/(SNR(ω)+1), (4)
and is widely used to reconstruct features of signals offline.
To obtain the system response h, and characterize the SNR of the signal of interest, we developed this protocol: the system response is obtained as a template, the temporal aligned mean of output signals produced by an impulse; the SNR is obtained as the ratio between the Power Spectral Density (PSD) of the template and the mean PSD of periodical acquisitions which do not contain signal of interest. This procedure can be considered as a calibration of the acquisition system that needs to be done periodically and does not impact on algorithm performance and resources utilization.
Still, the Wiener Deconvolution remains a computational demanding algorithm and is hardly usable in real-time, especially on an FPGA. The solution we present exploits the filter-like nature of the algorithm and applies it to signals as FIR filters: to simplify the problem, the algorithm has been divided into two sub-filters, one implementing the Wiener filter and one implementing Deconvolution. This way we can tune the two FIR filters independently to obtain the best possible solution.
The main application studied is based on the Jiangmen Underground Neutrino Observatory (JUNO) photo detection system (Photo-Multipliers Tubes (PMTs)) and its real-time signal processing board, based on an AMD Kintex 7 FPGA. The algorithm is used to extract Photoelectrons (PEs) hit Time and Charge (T\Q) information from the PMTs signals. As shown in Figure 1 and Figure 2, the algorithm better identifies close in time PEs than a simpler real-time T\Q reconstruction algorithm (COTi) and has comparable performance (around 98% of correct PEs identifications) with the offline deconvolution algorithm. Moreover, as shown in Figures 3-4, the RTWD algorithm reconstructs the charge of the PEs with a negative bias of ~20% with respect to the Offline Deconvolution algorithm that can be easily computed and calibrated offline.
