TNT Digital Pulse Processor - IEEE Xplore

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 53, NO. 3, JUNE 2006

723

TNT Digital Pulse Processor L. Arnold, R. Baumann, E. Chambit, M. Filliger, C. Fuchs, C. Kieber, D. Klein, P. Medina, C. Parisel, M. Richer, C. Santos, and C. Weber

Abstract—We report on the development of Tracking Numerical Treatment (TNT) boards, which are the basic bricks of a more ambitious data acquisition system intended for online data acquisition and treatment in future European experiments. The module makes extensive use of field programmable gate arrays (FPGA) for signal deconvolution and energy calculation with a minimum of data loss. Four channels can be processed simultaneously. Event collection is performed through a fast universal serial bus (USB2). The system can sustain an event rate of 100 kHz without dead time. Here we describe the practical implementation of pulse processing algorithms in a digital electronic module based on high density FPGAs for event processing. The overall system and its architecture are portrayed, along with some technical characteristics. Finally, we present some results and future developments. Index Terms—Data acquisition, digital spectrometry, digitizer, FPGA, high counting rates, online deconvolution, online pulse processing.

I. INTRODUCTION NT is a digitizer board developed to meet the requirements of current data acquisition systems in nuclear physics. Low dead-time even at high counting rates and good energy and time resolution for a large range of energies are the main issues. It is primarily targeted at data processing from Germanium detectors. TNT relies on the use of a state of the art FPGA, powerful enough to implement a full digital spectrometer providing trigger and energy computation in real time. It can include further custom functionality as waveform capture for subsequent analysis. The work presented here focus on the choice and online implementation of existing algorithms to achieve the requisites of nowadays experiments. In this paper, we present an overview of the whole system. In Section III we describe the signal processing algorithms implemented in the FPGA and finally the two main operating modes. We also describe the performance of the card in two in going experiments.

T

II. OVERVIEW TNT is a NIM-based card providing digital spectrometry and waveform acquisition of the input signals from detector’s preamplifier. It features four acquisition channels operating in parallel and offers good high-resolution spectroscopy. Each channel combines a fourth order anti-aliasing Nyquist filter (cut-off frequency of 40 MHz), high sampling speed (100 MHz) and 14-bits resolution data bus by using the AD6645 Flash Analog to Digital Converter (FADC) from Analog Devices [1]. Manuscript received June 3, 2005; revised March 2, 2006. The authors are with the IPHC, BP 28, F67037 Strasbourg Cedex 2, France. Digital Object Identifier 10.1109/TNS.2006.873712

On-board computation provides real time operation. A Xilinx Virtex II, three million gates FPGA [2] implements pulse processing and hardware control. The Virtex II architecture is characterized by dedicated resources, such as shift registers and embedded multipliers that are well suited for development of fast digital signal processing (DSP) algorithms. The parallel data treatment capabilities of FPGAs enables a higher throughput and reduces the event processing time, thanks to a large amount of on-chip memory. All those features are widely exploited in the implementation of algorithms in need of high logic integration, extensive pipeline and First Input First Output (FIFO) memory. An additional Spartan II FPGA offers system reconfiguration, giving the possibility to update the main FPGA contents via the USB bus. An 8051 compatible micro controller (FX2, from Cypress [3]) is responsible for data readout and slow control. Transfer is performed through a 16 bits, 48 MHz universal serial bus (USB) interface with an observed maximum rate of 30–40 Mbytes/s. The number of TNT boards that can be used in parallel depends on the required performances, since the whole system must share the USB bandwidth. The acquisition software, TUC (TNT USB Control), is in charge of slow control and monitoring of the boards in addition to providing tools for data visualization and saving (Fig. 1). It has been conceived to manage any number of TNT modules, assisting the user in changing settings. Spying of the data and counting rate is possible thanks to its graphical user interface. In addition, TUC can handle the updating of the embedded functionalities running on the Virtex II FPGA. Written in java, it has been in use on Linux and Windows platforms. Finally, it should be noted that the communication protocol between TNT modules and TUC is accessible and hence, TUC can be replaced by any homemade software. III. DSP ALGORITHMS Two well known DSP algorithms have been implemented in the form of recursive finite impulse response (FIR) filters. Deconvolution of the preamplifier signal grants an accurate energy measure. A performing trigger is needed for precise event detection in a wide range of amplitudes. Here we describe its practical implementation taking advantage of some advanced features in the FPGA. A. Deconvolution Digitization of the preamplifier output signals allows Trapezoidal shaping [4], [5], replacing traditional analog Gaussian shaping. It provides a good compromise between counting rate and low dead time. In addition, it offers good energy resolution

0018-9499/$20.00 © 2006 IEEE

724


Fig. 1. TUC, graphical user interface to TNT.

and excellent baseline stability at high counting rates. The trapezoidal shaping compensates the ballistic deficit effect during the rise time. Finally, it can be conceived in a recursive way, which means that any output sample can be obtained from the previous one plus a limited number of input samples. This form is well suited for an online implementation on dedicated hardware ([6], chapter 4), as shown in Fig. 2. The design extensively uses the embedded high-speed multiplier blocks and the fast carry chains for arithmetic operations. The abundance of SRL16 shift registers, an optimized Virtex II resource, is a key feature too. TNT benefits from FPGAs ability to perform several tasks in parallel. The FADC output is processed continuously using a pipelined, fast architecture to generate a real time shaped pulse. A wide range of filter parameters is provided for accurate setup and greater flexibility. The rise time and flat top width adjustments can both be extended up to 10 s to fine-tune spectrometer performance. Peaking time goes from 100 ns to about 20 s, with adjustable flat top duration. Pole/zero cancellation perform automatic fall time detection and correction.

Fig. 2. Oscilloscope snapshot showing the deconvolution and subsequent trapezoidal shaping (signal 2) of the preamplifier output (signal 1).

ARNOLD et al.: TNT DIGITAL PULSE PROCESSOR

725

IV. WORKING MODES

Fig. 3. Trigger scheme. The signal from the preamplifier is differentiated by the delay-subtract unit (DS), and then fed onto a moving average (MA) FIR unit. This filter is composed of a series of a DS and an accumulator (ACC) unit. The output of this stage is compared to a threshold and the result is delayed and combined using a logical and. This avoids false triggers from occurring.

Baseline restoration (BLR) is carried out through an exponential average of the baseline, authorizing its minute adjustment for applications requiring wide dynamic rate counting. BLR allows a proper suppression of the pedestal of the energy filter. It also considerably improves the energy resolution by suppressing the negative influence of detector leakage currents on base line stability. The output of the pulse shaper and other test signals are routed to a 12 bits, 100 MHz. Digital to Analog Converter (DAC) for diagnostic purposes. B. Trigger Low energy threshold is attainable thanks to the following dedicated triggering scheme, specially designed for a digital, real time hardware implementation ([7] chapter 4, and [8] chapter 5). The first sub module, a delay-subtract unit, differentiates the digitized preamplifier signal to remove any offset. The resulting signal is then fed onto a moving average unit, which acts as a low pass filter, hence removing noise. Finally, the signal is compared to a threshold value (leading edge trigger). Delays from the delay-subtract unit can be adjusted up to 320 ns. In order to prevent false trigger from signal glitches, the trigger signal has to achieve a minimum width and therefore the trigger filter has to exceed the threshold for a minimum time. The design is shown in Fig. 3. C. Clocking Each of the TNT modules includes an LVDS port, permitting the synchronization of several modules. All waveform digitizers, triggers and time stamps are driven by the same periodic signal through this port in a daisy chain configuration. TNT benefits of the advanced clock management capabilities provided by the AD9852 Direct Digital Synthesis (DDS) device [9]. The latter provides a stable, low jitter, high frequency resolution clock signal to feed the FADCs in order to obtain the best possible signal to noise ratio (SNR). Several general-purpose I/O NIM interfaces are available to define coincidence/veto acquisition schemes. Different delayed, variable length time-windows of up to 650 s can be synthesized. It is then possible to acquire data in coincidence or opposition to some event. Finally, an external event reference can be used to resynchronize all time stamp counters. This is also useful for synchronization with other systems. A picture of the TNT module, along with an outline of its main components is shown in Fig. 4.

The working mode of TNT is the following: each channel accepts signals directly from a detector preamplifier. The signals are digitized in order to apply real time digital processing. Before this, the offset can be altered to take advantage of the full dynamic range and avoid saturation of the FADC. Triggering, pile-up inspection and filtering of the data stream are performed by the FPGA, as well as averaging and detection of peak amplitude (Fig. 5). Every time a pulse is detected, parallel signal processing executes and outputs event data. The data is then transmitted over the USB interface to TUC. The latter acts as an event collector and increments a 32k Multi Channel Analyzer (MCA) spectrum, generating a histogram. This operating mode offers the best bandwidth to data flow ratio. A. Oscilloscope Mode A TNT2 module can continuously acquire waveforms with 14-bit precision, which are sent into the digital pulse shaper. This data flow is also stored into a circular buffer, implemented in an external memory to the FPGA. This memory has a depth of 1024K entries corresponding to 10.4 ms worth of event waveform data, at sampling intervals of 10 ns. When an event trigger is received, the module can incorporate waveforms of arbitrary length into its output data stream for offline analysis. As in a digital oscilloscope, it is possible to record pre-trigger waveform data. B. Processing Mode With the aim of improving energy resolution, customized hardware in the FPGA determines the average value at the flat top of the deconvoluted preamplifier signal. As shown in Fig. 5, this average can be calculated using an arbitrary number of samples. This general strategy is intended for better noise filtering, and includes the particular case of a single sample. The readout takes place over the USB using a low level block transfer protocol. This implies some buffering in order to avoid data loss due to non continuous data readout. The flexibility of TNT modules lies in the FPGA, which provides 96 blocks of 18 Kbits, fully dual-port embedded memory for critical signal processing applications [2]. It is used as FIFO data buffers here, with independent management of the read and write pointers and fully synchronous and independent clock domains for the read and write ports. Count vectors provide visibility into the number of data currently in the FIFO. Fig. 6 shows how this flag can be useful for inspection of possible data loss depending on the counting rate. Information from previously analyzed pulses is stored into these FIFOs while the data flow is continuously processed. No dead time occurs at this stage, since nor storage neither further treatment of the data flow are necessary for event generation. The depth and length of the data buffers can be configured according to the required information, up to the FPGA’s full memory capacity. Buffers are continuously filled with energy value, time of trigger occurrence (48 bits time stamp) and event counter on an event-by-event basis, which allows an offline inspection of the number of events rejected. This configuration

726


Fig. 4. The TNT2 board.

Fig. 6. FIFO filling level with time. When the FIFO is full (high level), no further event storage is possible. The slope of the trace depends on the event counting rate. Events at different rates correspond to different slopes. Very high counting rates saturate the event collector. Fig. 5. Pile-up rejection and energy averaging on the flat top. Trace 1 represents impulsions from the preamplifier while trace 2 shows corresponding trapezoids. For each impulsion, there is a trigger (trace 3). The energy is computed as an averaged value of flat top samples (trace 4). The two first measures are rejected due to trapezoid overlapping.

allows a maximum depth of 16K events. In parallel, this information is read out from the FIFO at up to 30 Mbytes/s. Finally, waveform capture and energy processing modes may run separately or in parallel. It provides offline pulse shape analysis of the full trace. For instance, position location and particle discrimination can be studied.

V. RESULTS In order to study the capabilities of the digital system, the energy resolution of a small planar Germanium detector was investigated. We compare here TNT with the performances of a classic spectroscopy array. Conventional analog NIM electronics perform Gaussian shaping and sampling using a low frequency, 13 bits ADC. Time occupancy with this type of shaping is about six times the peaking time, . The results of the tests with analog electronics are 0.7 keV FWHM for the 60 keV 241Am peak and 2.23 keV

ARNOLD et al.: TNT DIGITAL PULSE PROCESSOR

727

TABLE I PERFORMANCE TEST WITH TNT

FWHM for the 1.33 MeV 60Co peak with a of 4.8 s. When using a of 9.6 s, we measured 0.6 keV FWHM for the 60 keV 241Am peak and 1.87 keV FWHM for the 1.33 MeV 60Co peak. In the last case, the pile-up rejection made dead time increase noticeably up to 20 %. The counting rate was in both cases equal to 4 kHz. Table I summarizes performances with digital electronics. The test shows excellent results both in term of energy resolution and dead time even at high counting rates. No data loss due to the USB is observed at the highest counting rate, and all event rejection is due to pile up. The energy resolution at high counting rates is mainly limited by the detector. The dependence of the energy resolution on the trapezoidal filter parameters was systematically measured. The best energy resolution was obtained for a peaking time equal to 2.2 s, while still achieving an acceptable percentage of data rejection. The SNR decreases with low energies (low amplitudes), and in this case the contribution of noise from the electronics becomes more important. It makes the resolution very sensitive to sampling inaccuracies, such as differential non linearity (DNL). Promising results ([7], chapter 4) on DNL correction should be considered in further tests. TNT has been successfully used by the GRACE group for several months: the aim was to measure (n, xn) reaction crosssection induced by a neutron beam at GELINA (IRMM Geel) ([6], chapter 3). The first concern was the separation time between the gamma flash and the fastest neutrons which is not greater than 2.5 s. The other one was the beam frequency (800Hz). So on one hand, one needs a fast way to compute the energy within the 2.5 s time scale and on the other hand, one needs to overcome the acquisition dead time coming with the data readout in case of offline analysis. TNT’s digital, shorter pulse shaping time offers higher counting rates, and online treatment of the energies lowers the bandwidth. Our system has been tested in the measurement of the 207Pb(n,2n) reaction. For a 100% coaxial HPGe detector, the energy resolution on the 803 keV transition in 206Pb is equal to 2.6 keV, with a 3 s dead time. It allows the record of events corresponding to neutron energies up to 14 MeV, or to 3.9 keV with a 2.5 s, thus allowing the measurement of neutrons up to 20 MeV. The digital electronics bring a unique opportunity to perform spectroscopy of very heavy elements, where the productions cross section is less than afew 10 nb, with a high fission cross section. Very high counting rates are expected from this kind of experiments. In this case, online spectroscopy with low data rejection becomes a main concern to reduce data storage while

Fig. 7. Colbalt energy spectra taken at JYFL Jyvaskyla, in keV. The figure shows a zoom of the full spectra. The Compton suppression can be observed in semilog y axis.

still retaining good energy resolution. TNT2 cards will be used in 2 complementary projects: GABRIELLA at JINR DUBNA and JUROGAM II at JYFL Jyvaskyla [10]. For the latter project, the TNT cards will be linked to the existing TDR acquisition system where each detector’s information is associated to a timestamp given by a 100MHz clock (from METRONOME boards). Some tests regarding time alignment between the TNT cards and the TDR system have been successfully run recently at JYFL to assess its compatibility with current installation. The Compton suppression has also been tested with sources (Fig. 7) and with beam 36Ar on a107,109Ag target up to a 100 kHz rate per TNT channel. Analysis of the correspondingdata is under progress. Finally, the low noise of the analog stage, as well as the good performances in terms of SNR measures within the TNT make it a reference design for the digitizer part of the AGATA [11] collaboration. VI. FUTURE DEVELOPMENTS In order to improve the TNT system’s performances, an event collector and synchronization module is currently under development. It will distribute a common clock and time stamp reference through a 2 gigabits serial link, as well as some global commands, such as start acquisition and reset settings. The acquisition software TUC will be expanded and dispatched over the network allowing remote slow-control and monitoring. This will improve the whole system throughput since less computation will be done on the computers actually connected to the cards. This is in prevision of the use of more than 40 channels for experiments in Jyvaskyla and Dubna sites. ACKNOWLEDGMENT The authors would like to thank to thank the Grace group of IPHC, especially Dr. Gerard Rudolf and Dr. Strahinja Lukic for their helpful contributions in the development of algorithms. Tests for the future Jyvaskyla experiments were performed with

728


the help of Dr. Benoit Gall and Dr. Peter Jones. Our final thanks go to Dr. Gilbert Duchene for useful discussions and his support to this project.

REFERENCES [1] AD6645 flash analog to digital converter website, [Online]. Available: www.analog.com/en/prod/0,2877,AD6645,00.html [2] Virtex II field programmable gate array website, [Online]. Available: www.xilinx.com/products/silicon_solutions/fpgas/virtex/virtex_ii_ platform_fpgas/index.htm [3] Cypress semiconductor FX2 website, [Online]. Available: www.cypress.com/portal/server.pt?space=CommunityPage&control=SetCom munity&CommunityID=209&PageID=259&fid=14&rpn=CY7C680 13 [4] V. T. Jordanov and G. F. Knoll, “Digital synthesis of pulse shapes in real time for high resolution radiation spectroscopy,” Nucl. Instrum. Methods Phys. Res. A., pp. 337–345, 1994.

[5] V. Georgiev and W. Gast, “Digital pulse processing in high resolution, high throughput, gamma-ray spectroscopy,” IEEE Trans. Nucl. Sci., vol. 40, no. 4, pp. 770–779, Aug. 1993. [6] S. Lukic, “Mesure de sections efficaces de reactions (n, xn) par spectroscopie prompte aupres d’un faisceau atres haut flux instantane,” Ph.D. dissertation, Strasbourg, France, 2004 [Online]. Available: http://eprints-scd-ulp.u-strasbg.fr:8080/archive/00000245 [7] M. Lauer, “Digital Signal Processing for Segmented HPGe Detectors. Preprocessing Algorithms and Pulse Shape Analysis,” Ph.D. dissertation, Heidelberg, Germany, 2004 [Online]. Available: http://www. mpi-hd.mpg.de/cb/theses.html [8] L. Mihailescu, “Principles and Methods for A Ray Tracking With Large Volume Germanium Detectors,” Ph.D. dissertation, Bonn, Germany, 2000 [Online]. Available: http://www.fz-juelich.de/ikp/kernspektroskopie/luke/ [9] AD9852 Analog Devices DDS. [Online]. Available: http://www.analog. com/en/prod/0,C2877,CAD9852,C00.html [10] JUROGAM experiment website. [Online]. Available: http://www.phys. jyu.fi/research/gamma/jurogam/index.html [11] The Advanced Gamma-Tracking Array (AGATA). [Online]. Available: http://www-w2k.gsi.de/agata