Practical Implementations of Blind Demodulators

This paper examines the problem of demodulating time-dispersed digitally modulated signals with particular emphasis on two aspects, the all-digital implementation of such demodulators and the use of “blind” algorithms for initializing the demodulator in the absence of explicit training by the transmitter. Both were contrary to the directions taken in the original development of data communications systems but, paradoxically, both prove to be crucial to building the low-cost, bandwidth-efficient demodulators needed for modern commercial applications.

The classical demodulator for bandwidth-efficient QAM signals is “trained” by its counterpart and then, once operating properly, uses its own symbol decisions as the basis for continuously updating the demodulator’s various tracking loops. Over the past ten years there has been growing interest in the use of QAM demodulators which can acquire “blindly” as well, that is, which need no explicit cooperation from the transmitter to establish the communications link. Historically, interest in blind acquisition was limited to two related applications, non-invasive testing of equipment used for data transmission systems and signals intercept. In both cases it was necessary to acquire the transmitter’s signal without interference to its normal operation or to that of the intended receiver. Note that in both of these cases the transmitter and intended receiver still operate in a point-to-point fashion and will still use training to initialize each other. The monitoring receiver, however, must be capable of training itself without the transmitter’s help since the transmitter will send the training signal only when required by the intended receiver. The noninvasive test/intercept scenario is what motivated the blind equalization work reported in [1] and the fully blind QAM demodulator described in [2]. Foschini’s work [3] was motivated by the desire to restore the operation of a point-to-point digital microwave radio link after a fade without the need for cooperation from the transmitting terminal. While the necessary feedback path was available, the overall complexity of the transmission system was substantially reduced by not requiring its use.

Modern interest in blind equalization has been spurred by a different problem, the desire to operate digital point-to-multipoint and broadcast networks. The former of these inspired Godard’s suggestion of dispersion-directed blind equalization [4] in the late 1970s and Jablon’s work on a fully blind QAM demodulator [5] in the late 1980s. The broadcast application, an obvious extension of the multipoint problem, has now become practical for a variety of entertainment services, such as high-definition television (HDTV) and digital cable television. In these cases the need for blind acquisition stems from the desire to permit the transmitter to send its content unimpeded as the various receivers of the content come and go from the network. It is impractical in the extreme to require the transmitter to pause “revenue” transmission to train each new “client”, since it both interrupts the transmission of content and it requires a backward link to the transmitter from each receiver to request the training. A more subtle method, that of continuously embedding training signals within the content (thereby allowing the receiver’s tracking loops to operate in the classical fashion, albeit slower) is also undesirable since it dilutes the transmission rate of the revenue-bearing content.

What is needed then is a receiver which requires no special consideration from the transmitter in order to establish all of its tracking loops. What would be desirable is that this blind acquisition be accomplished with no more complexity than a trained demodulator and that there be no degradation in either demodulated signal quality or acquisition time. Most, but not all, of this is possible.

While an equalized QAM demodulator must contain processing steps for amplitude, timing, carrier, and equalizer acquisition, the classical use of a prearranged training sequence permits a significant degree of flexibility in the ordering of those steps, even though the various processing steps are coupled. It might not appear to the casual reader that the desire for blind acquisition would change this, but it does. Careful design of the training signal permits uncoupled measurements to be made of many key parameters at once. Initialized with these reasonably accurate values, the nested tracking loops will operate properly. In a blind demodulator, each of the parameters must be estimated in the presence of the partial or complete uncertainty about the others. This problem strongly impacts both the required ordering of the processing steps and the algorithms that can be used for each. An indication of this case can be seen by examining Table 1 which lists the first-order dependencies among various parameters to be tracked and others about which the demodulator will be uncertain. Note that all four are related to at least some of the other three.

Table 1. Interdependency Between the Various Demodulator Tracking Parameters

While other approaches are possible, practical experience, [2] for example, has shown the following processing order and choice of algorithms to work quite well:

Some of the following considerations go into this particular sequencing and algorithm choice.

Taking into account the observations made in the previous section, a generic QAM demodulator can be refined into the blind demodulator shown in Figure 1.

Figure 1. The Block Diagram of a Generic Blind Demodulator for PSK and QAM Signals

The operation of the demodulator shown in Figure 1 can be illustrated by examining the complex-valued signal at various points in the demodulator. The example in this case is a 30 Mbaud, 64-QAM digital microwave radio signal which has been received through an actual microwave channel. Noise has been added at a level of –26.5 dB relative to the received signal power. The example demodulator employs a 32-tap, T/2-spaced adaptive equalizer.

The signal is first power-adjusted, digitized, and then quadrature-downconverted to a complex-valued baseband representation. At this point the signal has not been sampled synchronously with the symbol rate and still contains a residual carrier frequency component. In addition, it is corrupted by channel distortions and additive noise. The “constellation” of this signal, shown in Figure 2(a), exhibits no obvious features of 64-QAM. The next processing step is the recovery of the symbol frequency and the resampling of the complex-valued signal to exactly two samples per symbol. The baseband constellation at this point, shown in Figure 2(b), still shows no features of a bauded waveform—this owing to the lack of carrier lock, the channel distortion, and the remaining additive noise. However, progress has been made, even if it is not visible. Since the signal is now synchronously sampled, it can be blindly equalized. In this case the CMA algorithm is used. This processing will cause the equalizer filter to (1) remove channel distortions and out-of-band additive noise and (2) refine the sampling phase of the signal. The output of the blind adaptation step is shown in Figure 2(c). The constellation now has a distinct ringed characteristic of a 64-QAM signal which is rotating because of a residual carrier term. Recall that CMA is carrier-phase-invariant and can therefore adjust the equalizer taps without the need for accurate carrier removal.

Figure 2. Constellation of a 64-QAM Signal as it is Acquired by a Blind Demodulator

The demodulator must now remove the residual carrier term. This is achieved using the “four-corners” technique. This method treats a high-order, square QAM signal as a much simpler QPSK signal by the strategy of only updating the carrier tracking loop’s estimate of the carrier phase when the instantaneous amplitude of the signal is big enough. By setting this amplitude threshold just inside of the four corner points of the constellation, only those corner points are used in the blind carrier acquisition. Figure 2(d) shows the “despun” constellation after CMA-based equalization and “four-corners” carrier recovery.

Once the signal has been equalized and the carrier component has been removed, the 64-QAM constellation points are clearly discernable in the baseband data. The carrier tracking and equalizer control loops may now be switched into the “decision-directed” mode, whereby the loop corrections are derived from the error between the baseband signal and the nearest, ideal 64-QAM constellation point. Following this step, the demodulator acquisition is complete. The constellation resulting from this step is shown in Figure 2(e). The demodulated signal’s SNR for this example is within about 1 dB of the ideal value of 26.5 dB, as established by the input SNR.

Figures 3, 4, and 5 show examples of practical blind demodulators. Figure 3 demodulates up to 24 voiceband modems. Figure 4 is a flexible demodulator for dispersive microwave channels. The demodulator shown there is fully digital and operates at symbol rates in excess of 40 MHz. The ASIC shown in Figure 5 is used for demodulating digital cable television signals. All of these implementations use the architecture described in this paper.

Figure 3. A Circuit Card Using Eight Texas Instruments TMS320C50 DSPs to Demodulate 24 V.33 128-QAM Modems Signals

Figure 4. A Single-Card Equalized Demodulator for 128-QAM Signals of Up to 40 Megasymbols/s

Figure 5. A Single-Chip Demodulator for 64- and 256-QAM Digital Television Signals

Twelve years ago knowledgable experts believed it to be a practical impossibility to blindly acquire and subsequently demodulate heavy dispersed, high-order QAM signals. Since then the required techniques have been developed, reduced to practice, and evolved through several cycles of implementation improvements. The technology has now entered the commodity world of commercial digital broadcasting. That success does not mean that all of the analytical questions have been answered, however, nor that there is not substantial opportunity for improvement.