See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224241473

VHDL implementation of feature-extraction algorithm for the PANDA

electromagnetic calorimeter

Conference Paper  in  IEEE Nuclear Science Symposium conference record. Nuclear Science Symposium · December 2010
DOI: 10.1109/NSSMIC.2010.5873866 · Source: IEEE Xplore

CITATION READS
1 282

5 authors, including:

Myroslav Kavatsyuk Elmaddin Guliyev

University of Groningen MINES ParisTech
187 PUBLICATIONS   3,292 CITATIONS    15 PUBLICATIONS   305 CITATIONS   

SEE PROFILE SEE PROFILE

Herbert Loehner Ganesh Tambave

University of Groningen University of Groningen
662 PUBLICATIONS   14,324 CITATIONS    16 PUBLICATIONS   165 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Hot hadronic matter View project

High energy physics View project

All content following this page was uploaded by Myroslav Kavatsyuk on 22 May 2014.

The user has requested enhancement of the downloaded file.

 VHDL Implementation of Feature-Extraction
Algorithm for the PANDA Electromagnetic
Calorimeter
M. Kavatsyuk, E. Guliyev, P.J.J. Lemmens, H. Löhner, Member, IEEE, G. Tambave
KVI, University of Groningen, Groningen, The Netherlands
on behalf of the PANDA Collaboration

Abstract—The feature-extraction algorithm, developed for the 200

digital front-end electronics of the electromagnetic calorimeter of

Signal level (V)

the PANDA detector at the future FAIR facility, is implemented in 150 APFEL
VHDL for a commercial 16 bit 100 MHz sampling ADC. The use LNP
of modified firmware with the running on-line data-processing
algorithm will allow to perform realistic performance studies of 100
the calorimeter and test a trigger-less readout concept of the data
acquisition for the PANDA experiment. 50

I. I NTRODUCTION 0
0 0.2 0.4 0.6 0.8
The mass of the proton and, therefore, of the visible
Time (μs)
Universe, originates from the strong interaction among the
three elementary light quarks confined in the proton, but Fig. 1. The pulse shapes measured for the discrete preamplifier (blue solid
the exact mechanism of the underlying dynamics is yet un- line) and for the ASIC APFEL-II with high gain output (red dashed line).
known. Charm-meson resonances and yet undiscovered glue-
balls might reveal the origin of the hadronic mass spectrum.
The PANDA collaboration at the future FAIR synchrotron expected in the barrel part of the EMC, with rates up to
facility at Darmstadt, Germany, will employ antiproton an- 150 kHz in the most forward region. The forward endcap
nihilations to investigate resonances in the charmonium mass EMC will be exposed to event rates up to 750 kHz. Due to
region. these rate variations, different photo-sensors and preamplifiers
The Electromagnetic Calorimeter (EMC) [1] of the PANDA will be employed for different parts of the EMC. For optimal
detector comprises of the central target calorimeter, read out by conditions of event selection the PANDA experiment will
large-area avalanche photodiodes (LAAPD), and the forward use a trigger-less data acquisition system. Each sub-detector
endcap EMC, read out by vacuum phototriodes (VPT). Cooled will continuously provide all single-hit event information.
PbWO4 (PWO) scintillating crystals will be employed for Therefore, the EMC-preamplifier signals will be continuously
the detection of high-energy photons, electrons and neutral digitized by sampling ADCs (SADC) and the data will be
mesons. The barrel EMC will be placed inside the 2 T processed on-line in FPGAs.
solenoid magnet of the target spectrometer. The gain of the The aim of the investigations presented here is to implement
selected photo-sensors is not sufficient to directly digitize in VHDL and test a simple, efficient, and robust feature-
the output signals. Therefore, special low-power and low- extraction algorithm, described in ref. [4], to perform the on-
noise preamplifiers were developed. The discrete component line signal-trace processing.
preamplifier (LNP) [2] will be used for the VPT readout,
and the ASIC APFEL-II [3] for the LAAPD readout. The II. M ATERIALS AND METHODS
ASIC has a built-in two-stage shaper and provides two output
The implementation of the feature-extraction algorithm [4]
signals with high and low gains. The discrete preamplifier is a
is done using portable VHDL and is tested on a Xilinx Spar-
one-range resistor-reset type with decay constant of 25 𝜇s.
tan 3 FPGA. To test the implementation and the performance
The signal shapes provided by the ASIC and the discrete
of the algorithm a series of measurements with the PANDA-
preamplifiers are shown in figure 1.
EMC prototype was performed. The test set-up consisted of
The PANDA experiment will make use of an antiproton
a single (20 × 20) × 200 mm3 PWO scintillating crystal with
beam on a frozen hydrogen-pellet target. Therefore, we expect
LAAPD photo-sensors attached at both (20 × 20) mm2 ends
a strong angular dependence of the event-rate distribution in
of the crystal. Both, LNP and APFEL preamplifiers were used
the EMC and, due to pellet density and frequency fluctuations,
to acquire data samples. The output signals were digitized by
also event-by-event fluctuations. Event rates of 10 kHz are
a 100 MHz 16 bit SIS3302 SADC [5]. For each event 10 𝜇s
Manuscript received November 12, 2010. long traces were stored for the off-line analysis. The obtained
 Time resolution, FPGA (ns)
Pulse amplitude, FPGA (V)

0.4
Y = 1.005 X Y = 1.02 X LNP Raw
3
0.1 MWD

Signal Level (V)

0.3 2

0.2 1 0
CFD
0.2 0.3 0.4
0
0 1 2 3 0.1
Pulse amplitude, software (V) Time resolution, software (ns) Timing

Fig. 2. The correlation of the measured pulse amplitude and the time 0
resolution obtained by analysing the same data using the software and the
hardware (FPGA) implementations of the feature-extraction algorithm. 0 0.2 0.4 0.6
Time (μs)

Fig. 4. The raw LNP signal and the resulting MWD pulses (upper panel, red
solid and blue dashed lines, respectively) and the CFD signal (lower panel).

applied. The raw data from the SADC is processed using

the Moving Window Deconvolution filter [6] (MWD). The
effect of the MWD filter on a LNP pulse is shown in figure 4.
The MWD allows to recover the exponential decay of the
LNP signal and produces a step-like function which, after
differentiation, results in a semi-rectangular shape of the
Fig. 3. Block-diagram of the feature-extraction algorithm implemented in required length. After the MWD shaping a constant signal-
VHDL.
offset is removed by the base-line follower. The base-line is
calculated as an average of a specified number of samples
data were analysed off-line using a PC and semi-online in before the pulse. Data of the detected pulses are not taken for
the FPGA. The FPGA analysis was performed using a Xilinx the base-line determination.
Spartan 3A-N design board. The transfer of measured data to The base-line subtracted signal is used for triggering and
the FPGA board and the read-out of the extracted time and time-stamp determination. The digital implementation of the
energy information were done using a serial RS232 interface. constant-fraction discrimination (CFD) algorithm is applied
Figure 2 shows the comparison of the results obtained using to obtain precise timing information. The linear interpolation
the software and the hardware processing of the same data. The using two data-points, see figure 4, is employed for finding
data, employed for the presented comparison, was measured the zero-crossing point with a much higher precision than
using above described single-crystal set-up with the APFEL the sampling period. For the EMC prototype it was possible
ASIC preamplifier. Correlation coefficients of about 99.8 % to achieve a time resolution better than 1 ns while using a
were determined. Similar results were obtained with the LNP 50 MHz sampling rate for the SADC.
preamplifier. The Moving Average filter (MA) is used to reduce high-
After verification of the VHDL implementation with the frequency noise. The pulse-detection is done employing com-
test data, the firmware was ported to the 100 MHz 16 bit bined information from the CFD and MA filters. Once the
SIS3302 SADC [5] module, which is used for the test experi- zero-crossing transition is detected in the CFD signal the
ments with the PANDA-EMC prototypes. The use of modified output of the MA filter is compared with a threshold. After
firmware with the running on-line feature-extraction algorithm the identification of the pulse the maximum value of the MA
will allow us to perform realistic performance studies of the signal is taken as an amplitude measurement.
calorimeter. Moreover, it will be possible to test a trigger- In case of a signal with finite rise-time, the MWD filtering
less readout concept of the data acquisition, namely to store provides a pulse with the required width, with non-distorted
information about all hits in all the channels of the calorimeter leading edge and the same length for trailing and leading
and identify cluster events using recorded time-stamps. edge, see figure 5. For the PANDA-EMC prototype equipped
with the LNP preamplifiers we found that the raw signal
III. F EATURE - EXTRACTION ALGORITHM has two components with different rise times. Therefore, to
The block-diagram of the implemented feature-extraction achieve the best energy resolution and lowest signal-detection
algorithm is shown in figure 3. Depending on the type of the threshold, it is necessary to use the MWD with a differentiation
input signal it is possible to select different processing paths constant longer than 200 ns, as shown in figure 5. The shorter
of the digitized data. For the LNP preamplifier, a resistor- differentiation constant reduces the resulting pulse amplitude.
reset charge-integrating preamplifier with the discharge time- This effect does not allow to get the optimal performance of
constant of 25 𝜇s, an additional digital pulse-shaping is the detector at high hit-rates as the long filtered pulses will pile
 200 Differentiation 80 ns Differentiation 200 ns
the whole duration of the pulse. The implementation of the
base-line follower foresees continuous averaging, as described
Raw Signal above. However, it is possible to force to perform the base-line
150 Single MWD initialisation only once. In this case, the same value is used
Amplitude (mV)

Double MWD
over all the running period.
100 The constant-fraction pulse is derived from the shaped and
base-line subtracted signal. The parameters of the CFD filter
are a fraction value and a delay. The delay should be adjusted
50 to match the rise time of the measured pulse. The maximum
possible value of the CFD-delay register is 32 SADC samples,
which corresponds to 320 ns at 100 MHz sampling rate. The
0 fraction value can be set to either 1/2 or 1/4. For the time-
0 0.2 0.4 0.6 0 0.2 0.4 0.6 stamping of the pulse the zero-crossing is detected. To achieve
Time (μs)
a sub-sample precision the successive interpolation between
the two samples is used, below and above zero level. The
Fig. 5. The pulse shape measured for the EMC prototype with LNP
preamplifier (black solid line) and resulting pulses after single (red dashed
interpolation is done with nine iterations yielding a precision
line) and double (blue dash-dotted line) MWD filtering. of 𝑇 /29 , where 𝑇 corresponds to the sampling period.
The energy readout is done after the MA smoothing, see
figure 3. The length of the MA filter can be adjusted by setting
up and such events will be lost. However, the use of the second the register values. After the pulse detection, the extracted
MWD filter allows to overcome this limitation. As can be seen energy and time values are written into the buffer memory.
in figure 5, the second MWD filter restores the original pulse The available buffer may store up to 4 million detected events.
amplitude even at shorter MWD differentiation constants at the Once the buffer is full, the new data will be replacing the oldest
cost of distortion of the rising edge of the pulse. Therefore, one. The acquisition can be stopped using either an external
to use all advantages of a double MWD filtering the second logic pulse, provided via the front panel of the module, or
MWD filter is employed only in the triggering and energy- by sending a special ”stop” command via VME bus. Once
readout chain, see figure 3. the acquisition is stopped, it is possible to read the data from
The above described feature-extraction algorithm is used in the buffer. The number of acquired events is provided by a
case of an unshaped input signal. If an analogue shaping is dedicated register.
employed, like in case of the APFEL ASIC preamplifier (see In addition to the described mode of operation, the firmware
figure 1), both MWD filters of the signal-processing algorithm has a number of debugging options. Namely, it is possible to
are not needed and, therefore, can be bypassed. store the output of all filters directly into the buffer memory.
This feature simplifies the procedure of optimizing the filter
IV. VHDL IMPLEMETATION parameters.

The used SIS3302 VME module [5] is a 8-channel 16 bit

V. O UTLOOK
100 MHz sampling ADC. Data from SADC chips are fed
to FPGAs for processing. For every two input channels one The aim of this project is to develop and implement in
Xilinx Spartan XC3s1000-5 FPGA is available. The incoming VHDL a simple, efficient, and robust feature-extraction al-
data from the SADC chips are 16 bit wide and of unsigned- gorithm for the front-end electronics of the PANDA EMC.
integer type. For processing with the feature-extraction algo- The above described implementation of the algorithm is a first
rithm this data are converted into 17 bit signed format and all step towards building a complete trigger-less readout chain.
further calculations are done with this precision. While the dedicated digitizer modules for the PANDA EMC
The MWD I and II filters, see figure 3, have two parameters, are being developed, the SIS3302 VME SADC was selected to
namely the length and the correction constant. Both of these perform test measurements with different detector prototypes.
parameters can be changed by programming corresponding The modified firmware of the SADC allows to process on-
registers of the SADC module. In order to save FPGA re- line digitized wave forms and to store only the energy and
sources, the MWD-length parameter can take values which time information of the detected hits. This data is transferred
are powers of two: 2, 4, 8, 16, 32, 64. This selection of the to the PC, where the search for hit correlations takes place.
MWD length simplifies the implementation of the division, Simultaneous hits in different channels are combined into
one of the internal operations required by the filter. Both MWD events.
filters in the feature-extraction algorithm can be bypassed by The current version of the feature-extraction algorithm is
setting register values. Such a possibility allows to use the very flexible in terms of parameter selections. This allows to
same firmware for the shaped and non-shaped signals. achieve the best performance with different types of signal
The base-line follower determines and removes the constant sources. Therefore, this development might be of interest for
offset of the incoming signal. To determine the base-line an other experiments. For the PANDA EMC the aim is to find the
average of 512 data-samples before the pulse is used. Once best combination of the parameters. With the current design of
the pulse is detected, the base-line averaging is deactivated for the firmware it is possible to process data for 16 channels with
the second smallest Xilinx Virtex6 FPGA, leaving room for the
implementation of data-transfer protocols. For the final design
of the feature-extraction algorithm most of the parameters
will be hard-coded in the firmware, therefore, even more
reducing resource requirements. In addition to the described
functionality, a pile-up detection and recovery algorithm will
be implemented to improve the performance of the detector at
high counting rates.

R EFERENCES
[1] PANDA EMC Technical Design Report, arXiv:0810.1216v1.
[2] W. Erni, M. Steinacher, Univ. Basel, in PANDA Technical Progress Report
(web: http://www-panda.gsi.de/framework/documents.php), Feb. 2005.
[3] P. Wieczorek, H. Flemming, GSI report (2007) 30.
[4] M.Kavatsyuk et al., Conference Record, 2009 IEEE Nuclear Science
Symposium, Orlando, Florida, USA, October 25-31 2009, N10-4.
[5] SIS3302 SADC, http://www.struck.de/sis3302.htm.
[6] A. Georgiev, W. Gast, IEEE Trans. Nucl. Sci. NS-40 (1993) 770; J. Stein
et al., Nucl. Instr. Meth. B 113 (1994) 141.

View publication stats