J.D. Behm, Department of Defense(1); T.J. Endres, Sarnoff Digital Comm. (2); P. Schniter, Cornell University (2); C.R. Johnson, Jr., Cornell University (2);
C. Prettie, Applied Signal Technology; M.L. Alberi, ETIS; I. Fijalko, ETIS

• Abstract
• Introduction
• Data Collection
• Data Processing
• Data Categorization
• Observations
• Conclusion
• References
• Footnote

Abstract

This paper reports on the gathering, processing, and categorization of empirically derived time-varying channel responses. The passband data and data collection information is provided courtesy of Applied Signal Technology (Sunnyvale, CA). It is the intent of this paper to provide the signal processing community with a database of time-varying fractionally-spaced channel responses and received sequences based on empirical measurements which can be used to test and refine existing time-varying channel models and also propose new ones.

Back to top of page

Introduction

A common assumption in analysis of blind equalization and identification techniques is that of a linear, time-invariant channel model. Many existing and emerging applications, however, challenge this time-invariant assumption. While many time-varying models are proposed in the literature, some are suspect in a practical setting and few are data-based.

In the Spring of 1996, Applied Signal Technology (Sunnyvale, CA) collected data to empirically assess the impact of a wideband mobile communication environment on digital communications [1]. A vehicle with a receiver and an antenna collected digital microwave transmissions from stationary sources for approximately six weeks in Northern California. We at Cornell University and our colleagues were given access to Applied Signal Technology’s raw field data with the promise to “prepare” it for use by the general signal processing community.

Our intent with this data is to provide the community with an empirically-derived database which can be used to test and refine existing time-varying models and possibly propose others in an effort to meet the needs of today’s demanding applications. To this end we have written demodulation software (MATLAB and C) which provides, among other things, (approximately) length-130,000 T/2-spaced, complex-baseband received sequences, and successive channel estimates over this observation window. Moreover, due to Applied Signal Technology’s substantial effort in the field, the database is quite large. Hence, we also attempt a classification of the data into three (possibly overlapping) categories: stationary or slowly time-varying, non-stationary, and unprocessable using standard blind demodulation techniques such as CMA [5].

The sequel is organized as follows. Data Collection describes the data collection procedure and field experiments. Data Processing describes our subsequent data processing and demodulation procedure. Data Categorization provides some demodulation results and a classification of the experiments. Observations lists some observations based on the data and Conclusions contains concluding remarks and a summary of Internet addresses for data access.

Back to top of page

Data Collection

In April and May of 1996, Applied Signal Technology performed experiments in the Northern California area near Red Bluff to determine demodulation requirements for on-the-move (OTM) high data rate digital communications. A fixed source with 40 MHz bandwidth at radio frequencies of 4.45 and 7 GHz transmitted QPSK data at 50 Mbps (25 Mbaud). The source, although stationary during a single experiment, was moved several times during the six week period. The receiver was battery powered and mounted in a four-wheel-drive vehicle with a horn antenna above the roof. OTM data was collected for mobile velocities from 5 to 50 mph at distances between 1 and 40 miles. We calculate the impact of the Doppler shift on the signalling rate to be less than 2 Hz or approximately 1% of the observed baud frequency timing error. The physical characteristics of the experiments varied greatly, from having an unobstructed line of sight to being shadowed by a hill or being blocked by a passing truck.

The data was collected using Applied Signal Technology’s Model 195 Snapshot Recorder/Analyzer with 64 MBytes of memory and a sample rate of 200 MHz, which at 25 Msymbol/sec corresponds to 8 samples per symbol.(3) A sample power spectrum of the 70 MHz IF receiver output is shown in Figure 1. Typically, OTM data was collected in 0.5 MByte successive snapshots at 0.1–0.5 second timer controlled intervals and stored on disk for subsequent processing. It was predominantly these OTM multi-snapshots of data collection, separated by off-the-air intervals, that we post-processed in Data Processing. There exist 114 data files, most of which contain 8–40 0.5 MByte successive snapshots, for a total of 1.2 GBytes of data representing varied physical experiments.

Figure 1. Passband spectrum

Back to top of page

Data Processing

The data processing is all MATLAB software based (with C-MEX files) and consists of two primary functions; a QPSK demodulator and a channel estimator.

The QPSK demodulation software is comprised of four blocks (see Figure 2):

1. a data reader and converter: getdata(·)
2. a complex baseband/resampler: c_base_resamp(·)
3. a looping blind equalizer: cmaloop(·)
4. a looping DD carrier-tracker/equalizer: ctrac_eql(·).

Figure 2. Demodulation software flow

Block 1 simply reads in the packed binary data and converts it to an array of floating point values. Block 2 nominally complex basebands the signal and uses band-edge timing recovery (BETR) [4] and (interpolated-coefficient) polyphase resampling to provide baud-synchronous sampling at twice the symbol clock rate. The BETR technique is aided by an interpolated-FFT technique which estimates the mean deviation in received symbol rate from the specified 8 samples per symbol. Block 3 takes the T/2-sampled output of Block 2 and blind equalizes using the Constant Modulus Algorithm (CMA) [5] over the first half of a single snapshot. The module makes multiple (3 was the number used in the processing reported here) forward and backward passes, maintaining baud continuity, to reduce the error rate sufficiently for transfer to a decision-directed (DD) equalization mode. Block 4 uses the equalizer estimate from Block 3 and simultaneously does equalization and decision-directed carrier tracking. As in Block 3, the software makes multiple (2 was used in the processing) forward/backward phase-continuous passes through the data. The primary outputs are the numerically controlled oscillator (NCO) values representing the residual carrier, and soft and hard symbol decisions. The NCO is applied to the nominally complex basebanded data (frequency translated down by 70 MHz) to remove the residual carrier and hence provide the “desired output” sequence for the channel estimator.

The channel estimator uses LMS [3] (RLS was also tried with similar results) to provide estimates of the complex baseband T/2-spaced channel. (See Figure 3.) The hard decisions (estimated symbols) with interleaved zeros comprise the “input” sequence and the complex basebanded data comprises the “desired output” sequence. (See [2] for further discussion of this channel estimation procedure.) The two sequences are complex correlated to determine an appropriate system delay and then the LMS algorithm is run over successive sections (most of the experiments use 4 sections, each which are 1/4 of the 0.5 megasample snapshot) of the data snapshot.

Figure 3. Channel estimation software flow

Both the equalizer and channel filters were 100 coefficients, which appeared to be adequate for almost all files. A step size of .001 was used for both CMA and LMS. Some experiments were conducted with reducing the step size for successive loops of the blind equalizer but there was no significant change in the quality of the demodulation.

All of the software was written to run under MATLAB 4.2. The algorithms used in the demodulation are all designed for arbitrary QAM signals. However, for expediency and efficiency some of the functions and scripts have been tailored to QPSK. The equalizer, carrier tracker and LMS routines have been re-written in C and compiled as MATLAB executables known as MEX functions. These MEX functions run 10 to 50 times faster than the corresponding MATLAB functions and allowed processing the entire 1.2 Gigabyte database in a reasonable amount of time.

Back to top of page

Data Categorization

With such a large amount of field data gathered, a useful task was to delineate the severity of the signalling environments for the various experiments and group the data files. To this end, using the demodulation software described in Data Processing, we selected OTM files and categorized a subset of them. Presently, selected data (in the form of passband data, channel estimates and T/2-spaced received sequences) and the demodulation code discussed in Data Processing are available at Cornell University BERG’s web page though the intent is to move this information to the signal processing database maintained at Rice University [6].

We looked for short-term (within the time typical for convergence of the blind equalization algorithm) time variations. Our method for categorization therefore considered four consecutive channel estimates over a snapshot (i.e., one channel estimate every 130,000/4 T/2-spaced observations) and determined (with the help of the measures returned from the software) if these channel estimates suggested (i) legitimate time variations, (ii) nearly stationary environments, or (iii) poor demodulation. Tables 1, 2, and 3 show our classification of files according to these categories, respectively. Because each file contains multiple snapshots (8–40 contiguous .5 Mbyte data blocks) we classified the entire file as having significant time variations (i.e., in (i)) if any of the single snapshots suggested this behavior. Indeed, it was often the case for those files which suggested legitimate time variations that only a handful of snapshots (out of the 8–40 possible) motivated the file’s inclusion in this category. In such cases, we marked which snapshots were of interest.

Table 1. Significant Time Variations

Table 2. Stationary or “Mild” Time Variations

Table 3. Demodulation Errors

Our intent was that this categorization aid in minimizing the initial work that would otherwise be necessary by other researchers in using this data. We admit, however, that our processing was not exhaustive and more experiments could be performed to optimize the demodulation of files for which the equalizer failed to “open the eye.”

Back to top of page

Observations

Our main observations based on the data include:

1. Many experiments suggested insignificant time variations or nearly stationary signalling environments after baud synchronous resampling. See Table 2.
2. The “significant” portion of the estimated channel impulse responses was typically within a 500 nsec window.
3. There were instances of significant multipath in which a secondary ray was nearly the same magnitude as the primary. For example, see Figure 4 which shows snapshot 8 of file hillshadow.otm.4GHz.
4. The majority of the files in Table 3 were likely low SNR files due to the receiver being over 30 miles from the transmitter.
5. In some cases there were large amplitude variations in a secondary ray. For example, see Figure 5 which demonstrates longer-term time variations by showing one channel estimate each from snapshots 7–10 of file hillshadow.otm.4GHz.
6. A lack of baud synchronization can be mistaken for a channel time variation, where, for instance, the estimated channel coefficients can be seen to “roll” in time. For example, see Figure 6 which is a “close-up” (shows channel taps 40–60) of snapshot 15 of file oxbox.otm. However, in most cases the baud-timing estimation was accurate enough so that no time variations attributable to timing errors were evident.
7. It was observed that the attenuation of the channel varied significantly over time for some of the files. For example, see Figure 7 of snapshot 13 of file bowman.4GHz.VV.otm.
8. In some instances, the software discussed in Data Processing was able to reliably demodulate data files for which the techniques of [1] failed.
9. The field data was created using a (hardware) degree 15 linear recursive bit generator, which unfortunately began malfunctioning, producing bit slips in much of the OTM data files. Thus, though the underlying structure was used for error estimates, it could not reliably be used for error correction.

Figure 4. Consecutive channel estimates from snapshot 8 of hillshadow.otm.4GHz showing a large secondary ray.

Figure 5. Channel estimates from snapshots 7–10 of hillshadow.otm.4GHz showing time variations spaced .2 seconds apart.

Figure 6. Consecutive channel estimates from snapshot 15 of oxbox.otm showing the effect of a baud frequency timing error.

Figure 7. Consecutive channel estimates from bowman.4GHz.VV.otm showing time varying attenuation.

Back to top of page

Conclusion

This paper has summarized the collaborative efforts of Applied Signal Technology and the Blind Equalization Research Group at Cornell University in providing an empirically-derived database to study the time-varying effects on digitally modulated signals. Our efforts are by no means exhaustive, and we invite other researchers’ comments and efforts regarding this data. Please notify us if you access and use the database in your research and development studies.

Back to top of page

References

1. Applied Signal Technology, “Results of Laboratory and Field Measurements of Propagation Effects for On-the-Move High Capacity Trunk Radio,” Scientific and Technical Report B001, November 1996.
2. R.P. Gooch and J.C. Harp, “Blind Channel Identification Using the Constant Modulus Adaptive Algorithm,” Proceedings of the International Conference on Communications, June 12–15, 1988.
3. S. Haykin, Adaptive Filter Theory, Englewood Cliffs NJ: Prentice Hall, second edition, 1991.
4. D.N. Godard, “Passband Timing Recovery in an All-Digital Modem Receiver,” IEEE Transactions on Communications, vol. 26, no. 5, pp. 517–23, May 1978.
5. C.R. Johnson, Jr., P. Schniter, T.J. Endres, J.D. Behm, R.A. Casas, D.R. Brown, and C.U. BERG, “Blind Equalization Using the Constant Modulus Criterion: A Review,” Proceedings of the IEEE. Invited for special issue on blind identification and equalization. Submitted July 1997.
6. D.H. Johnson and P.N. Shami, “The Signal Processing Information Base,” Signal Processing Magazine, vol. 10:36–42, October 1993.

Back to top of page

Footnotes

1. This work was performed while J.D. Behm was a visiting scientist at Cornell University and while T.J. Endres was a graduate student and post-doc at Cornell University, January–August 1997.
2. Supported in part by NSF Grant MIP-9509011 and Applied Signal Technology.
3. Our demodulation procedure in Section 3 resamples the data to two samples per symbol, as well as accounting for baud frequency errors which we observed to be on the order of 100–200 Hz.