The invention also relates to an apparatus according to the claim 6 suited for per- forming the training phase of adaptive channel equalization on a digital communia- tions path, whose data transfer channel has a stop band (or stop bands) on the signal spectrum, and further to a transmitter according to claim 11 and to a receiver according to claim 13.

In the text describing the prior art and the features of the present invention, the following abbreviations are used: CAP Carrierless amplitude and phase modulation DFE Decision-feedback equalizer FFE Feedforward equalizer LMS Least-mean-square PAM Pulse amplitude modulation QAMQuadrature amplitude modulation RX Receiver TX Transmitter TML Tomlinson-Harashima precoder Copper wirelines serving subscriber connections will be used for data transmission at increasingly higher speeds. Plausibly the transmission speeds will be extended to tens of megabits per second and higher. Herein, the bandwidth of the modulated signal will be in the order of 10 MHz.

When the modulation bandwidth reaches the megahertz range, different kinds of radio-frequency interference appear problematic particularly when overhead cables are used for subscriber connections. The modem receiver (RX) may also be subject to interference from radio stations emitting broadcast signals and radio-amateur stations, whose radio-frequency transmissions may be captured by the overhead cable so as to cause a so-called transverse interference signal component in the twisted pair cable and thus reach the receiver of the modem connection in an inter- ference-causing manner. In a converse manner, the modem connection may disturb radio station listeners or reception at a radio-amateur station, because the power spectrum of the modulated signal traveling along the pair cable contains energy at the radio-frequency bands mentioned above and a portion of this energy is radiated to the environment.

The radio-frequency interference caused by the transmitter (TX) of the modem con- nection can be eliminated by way of filtering the power spectrum of the modem output signal free from emissions at the frequency bands coinciding with those subject to interference. The bandstop filters required herein may be implemented using analog, digital or both of these techniques. The radio-frequency interference coupled over the cable to the modem receive circuit may be filtered away by provid- ing the receiver (RX) with suitable bandstop filters that may be implemented using analog, digital or both of these techniques, respectively.

In fast modems, signal processing generally is accomplished digitally, whereby the receive signal is first converted into digital format by an analog-to-digital converter (AD converter). Inasmuch the level of the radio-frequency interference emission may be high as compared with the received data signal, it is advantageous to in the opti- mum utilization of the dynamic range of the AD converter to perform the elimination of the radio-frequency interference at least up to a certain degree by means of band- stop filters that are realized with the help of the analog techniques and placed preced- ing the AD conversion.

On a modem connection, the signal transmission path (cable, line transformers,

filters, etc.) cause amplitude and phase distortion on the signal that is generally compensated for by means of equalizers that are located in the receiver, the trans- mitter or partially in both. Bandstop filters used in the receiver, the transmitter or in both of these communicating units are particularly problematic as to the signal distortion, because the filters form a signal-eliminating stop band (or stop bands) on the signal spectrum, thus blocking or essentially weakening the transmission of certain frequency components of the data signal, too. In the discussion to follow, the bandstop filters located in the receiver, the transmitter or partially in both are considered to form a portion of the transmission channel.

An equalization technique of channel distortion commonly used in modems based on a linear modulation system (PAM, QAM, CAP) is to employ an adaptive linear equalizer (FFE) and an adaptive feedback equalizer. If the feedback equalizer is located in the receiver, it is called a decision-feedback equalizer (DFE) and, respec- tively, if located in the transmitter, it is known as a Tomlinson-Harashima precoder (TML). The system may also be provided with both the DFE and the TML. Further- more, the linear equalizer may be located in the receiver, the transmitter or a part thereof in the transmitter and another part in the receiver.

In the following, a digital communications channel is examined in terms of the training phase of its adaptive equalizers. The line code used on the channel may be either a quadrature-amplitude modulation (QAM) or a carrierless amplitude and phase modulation (CAP). In FIG. 1 is shown a model for a system implemented using conventional techniques, wherein the receiver is provided with an adaptive linear equalizer (FFE) and an adaptive decision-feedback equalizer (DFE) (cf. Lee & Messerschmitt). The fixed filters and modulation schemes are included in the channel noise model (CHN). The outgoing bit stream is coded into symbols (S) that are sent through the channel 2. In the receiver, the output signal of the channel 2 is processed by equalizers (FFE and DFE), and the decisions on symbols (S') are made from the equalized signal. The decision resulting in the resolved symbol (S') is also called the estimated received symbol. The transmitted bit stream is recovered on the basis of the estimated received symbols (S') resolved by the receiver (RX). Both

adaptive equalizers are adapted to the characteristics of the channel 2 during the training period carried out when a connection is being established. The equalizers are also continually adjusted during the period of data transmission in order to compen- sate for possible changes in the channel 2. The equalizers are adapted and controlled on the basis of the detection error (e) of the receive signal with the help of a least- mean-squares (LMS) algorithm.

In FIG. 2 is shown another system according to the prior art (cf. Lee & Messer- schmitt). The receiver has an adaptive linear equalizer (FFE), while the transmitter has a feedback equalizer (of the TML type). During the training period, also this system operates in the same fashion as that illustrated in FIG. 1 using a linear equalizer (FFE) and a decision-feedback equalizer (DFE). At the end of the training period, the tap-weight values of the decision-feedback equalizer (DFE) are trans- mitted over an upstream auxiliary channel to the transmitter, wherein they are utilized in the configuration of a Tomlinson-Harashima precoder (TML).

The methods generally used for training an adaptive channel equalizer to the charac- teristics of a transmission channel are blind training, which takes place from data, and forced training. In blind training, the transmitter (TX) sends during the training period a similar signal as that of a normal data transmission state and the receiver (RX) has no a priori information on the sent symbols. Hence, blind training is based on the statistical properties of the receive signal, wherefrom the receiver can adjust the equalizers into a correct state (e. g., assuming that each symbol of the symbol con- stellation occurs at an equal statistical probability). Generally, it is necessary to use a dedicated training detector during the training period in the same fashion as a circle detector is used in certain types of voice-frequency modems, for instance. FIG. 3 shows the timing diagram of the functional steps of the timing diagram of blind training. In forced training, the receiver (RX) has a priori information on the data (i. e., the training sequence) being sent by the transmitter (TX) during the training period. The equalizer adjustment is based on the difference values between the a priori known training sequence and the received signal. A problem hampering forced training is that the transmitter (TX) must submit the receiver (RX) signaling informa-

tion on the start instant of the training sequence. Under the circumstances of practical channels, the transmission of such a start signal with a sufficiently high timing accuracy is often difficult. FIG. 4 shows a timing diagram of the functional steps of forced training.

In blind training the stop bands placed on the signal spectrum cause problems. To generate such stop bands, the transfer function of the communications channel must have one or more nulls that fall within the frequency spectrum of the data signal.

These spectral nulls deteriorate the transmission of certain frequency components of the data signal so much that in terms of channel equalization, it is more correct to discuss the nulls of the transmission channel frequency spectrum that entirely prevent the transmission of the data signal at its frequency components falling on the spectral nulls of the channel. A linear equalizer (FFE) is incapable of fully compensating for the distortion caused by the spectral nulls of the channel, because a complete compensation of a spectral null in the channel transfer function would require infinite gain in the transfer function of the linear equalizer (FFE) at the frequency of each spectral null of the transmission channel. In other words, when there are spectral nulls in the transfer function of the channel and only a linear equalizer (FFE) is available, the equalizer output will inevitably bear some degree of intersymbol interference (ISI). The degree of intersymbol interference in proportion to the signal energy at the output of the linear equalizer (FFE) is dependent on such factors as the number, frequencies and Q-factors of the spectral nulls. Intersymbol interference complicates correct decision-making on received symbols in the same manner as noise. On the other hand, a sufficiently high proportion of the symbol decisions should be correct in order to permit the decision-feedback equalizer (DFE) to adapt so as to perform the compensation of channel distortion. Hence, blind training fails if the channel transfer function has such spectral nulls that a linear equalizer (FFE) alone is incapable of giving a sufficiently good symbol error rate for the adaptation of the decision-feedback equalizer (DFE).

It is an object of the present invention to overcome the drawbacks of the above- described techniques and to provide an entirely novel type of method and apparatus

for use on a digital communications channel.

The goal of the invention is achieved by way of performing the training of adaptive equalizers on a channel having stop bands different from zero frequency on the frequency band of the data signal using a method which is otherwise similar to blind training but uses during the training period such a symbol constellation that a linear equalizer (FFE) alone gives a situation, wherein a sufficiently high proportion of the symbol decisions are correct for the purpose of adapting the decision-feedback equalizer (DFE) to the channel distortion. Then, the transmitter (TX) need not send the exact start instant of the training sequence to the receiver (RX).

More specifically, the method according to the invention is characterized by what is stated in the characterizing part of claim 1.

Furthermore, the apparatus according to the invention is characterized by what is stated in the characterizing part of claim 6.

The transmitter according to the invention is characterized by what is stated in the characterizing part of claim 11.

The receiver according to the invention is characterized by what is stated in the characterizing part of claim 13.

The invention offers significant benefits.

The invention makes it possible that even under a situation where the channel has stop bands on the frequency spectrum of the data signal, the training of adaptive filters to the channel distortion does not need the transmission of the exact start instant of the training sequence from the transmitter (TX) to the receiver (RX).

In the following, the invention is described in more detail with reference to exempli- fying embodiments elucidated in the appended drawings in which

FIG. 1 shows a block diagram of a system of the prior art for implementing channel equalization; FIG. 2 shows a block diagram of another system of the prior art for implementing channel equalization ; FIG. 3 shows a timing diagram of a conventional training process of adaptive equalizers to channel distortion; FIG. 4 shows a timing diagram of another conventional training process of adaptive equalizers to channel distortion; FIG. 5 shows two possible symbol constellations usable in QAM or CAP modulation schemes and also diagrammatically the eyes within which the output states of a linear equalizer fall due to intersymbol interference caused by stop bands located on the signal spectrum of the transmission channel; FIG. 6 shows a possible symbol constellations usable in QAM or CAP modulation schemes and, respectively, a sparsely mapped symbol constellation suitable for use in the method according to the invention; FIG. 7 shows a timing diagram of a training process according to the invention of adaptive equalizers to channel distortion; and FIG. 8 shows in greater detail a timing diagram of a training process according to the invention for adaptive equalizers to channel distortion.

Accordingly, the invention concerns a method and apparatus suitable for implement- ing the training of adaptive equalizers to channel distortion in a situation, wherein the channel includes bandstop filters that cause one or a greater number of stop bands on the transmission spectrum of a data signal. The theoretical basis of the method will

be evident to the reader from the discussion given below.

Referring to FIG. 5, therein is shown in the complex plane two possible symbol constellations (A and B) that are suited for use in a QAM or CAP modulation scheme, as well as diagrammatically outlined eyes (L) within which the output states of a linear equalizer fall due to intersymbol interference caused by stop bands located on the signal spectrum of the transmission channel. Symbol constellation A repre- sents a 4-QAM scheme (or, alternatively, a 4-CAP scheme), while symbol con- stellation B represents a 16-QAM scheme (or, alternatively, a 16-CAP scheme). The intersymbol distance between neighboring points of the symbol constellation is denoted by letter d.

Even if the channel happens to be entirely noise-free, the output states of a linear equalizer (FFE) would not map exactly on the points of the symbol constellation, but rather, the intersymbol interference caused by the stop bands would make the symbol states to scatter over the eyes (L) surrounding the ideal points of the symbol constellation. From FIG. 5 it is easy to see that the risk of incorrect symbol recovery decisions is essentially higher in symbol constellation B than in constellation A, since in constellation B the eyes (L) corresponding to the different symbol points overlap with each other.

The area of the eye (L) surrounding a symbol mapping point is proportional to the strength of the intersymbol interference. With the help of conventional signal analysis, it can be shown that the strength of the intersymbol interference in turn is proportional to the signal power.

The invention is based on the following concept: 1) The strength of the intersymbol interference due to channel stop bands at the output of the linear equalizer (FFE), indicated as the area of the eye (L) per symbol, is proportional to the signal power.

2) The ratio of the intersymbol distance (d) between neighboring points of the

symbol constellation to the signal power and, hence, to the area of the eye (L) can be increased by way of making the symbol constellation sparser.

A corollary of the above reasoning is that the proportion of incorrect symbol recov- ery decisions becomes smaller when the symbol constellation is made sparser.

The method according to the invention for training adaptive equalizers for the com- pensation of channel distortion is otherwise similar to conventional blind training with the exception that during the training period, there is used a sparser symbol constellation, whereby it is possible to reach even with a linear equalizer (FFE) alone a situation, wherein a sufficiently high proportion of the symbol decisions are correct, thus allowing the decision-feedback equalizer (DFE) to be trained for the compensation of channel distortion.

Referring to FIG. 6, therein is shown a possible symbol constellations A usable in QAM or CAP modulation schemes and, respectively, a sparsely mapped symbol constellation B suitable for use in the method according to the invention. The sparse constellation (B) is formed from the original symbol constellation (A) of the data signal states by selecting only the extreme corner points for use. It must be noted herein that generally the number of different symbol states (S) in a symbol constella- tion used during the training period is smaller that the number of symbol states in the constellation used during the actual data transmission. Advantageously in the spirit of the method according to the invention, the symbol points of the sparse symbol constellation used during the training period coincide with the corresponding points of the original symbol constellation as shown in FIG. 6. While this feature is not necessary under all circumstances, it generally makes it easier to the switch from the training phase to the data transmission state.

Obviously, the sparse symbol constellation may also be used for signaling between the transmitter (TX) and the receiver (RX).

The functional steps of the method according to the invention are elucidated in the

timing diagram shown in FIG. 7. Herein, transmitter (TX) sends data during the training period using the sparse symbol constellation. Receiver (RX) adjusts the equalizers using a similar (sparse) detector corresponding to the sparse symbol con- stellation. After the equalizers have been trained for a given time or a stable state has been reached, the system can be switched to data transmission with the provision that the error measurable at the detector input is sufficiently low. At the transmitter (TX), switching to the data transmission state also means adoption of the symbol constellation of the data transmission state, while the receiver (RX) in a respective manner switches from the sparse symbol constellation to a detector corresponding the symbol constellation of the data transmission state. In terms of the mutual timing between the transmitter (TX) and the receiver (RX), it is sufficient that the trans- mitter (TX) sends data using the sparse constellation for such a time that receiver (RX) can finish the training of its equalizers before the transmitter (TX) switches over to the data transmission state. In practice, this can be accomplished using predetermined state intervals.

A more detailed timing diagram of a possible training phase sequence accomplished using the method according to the invention is shown in FIG. 8. During the training period, transmitter (TX) sends data using the sparse symbol constellation whose points coincide with the respective points of the symbol constellation used in the data transmission state. In the beginning of training, receiver (RX) initializes the tap gains of the linear equalizer (FFE) and the decision-feedback equalizer (DFE) to zero. Sub- sequently, the gains of certain taps (1 to 10) in the linear equalizer (FFE) are adjusted with the help of an LMS algorithm using the sparse constellation. The taps being adjusted are selected from the group of taps where the main tap gain is desired to become largest. Next, all the tap gains of the linear equalizer (FFE) are adjusted using the sparse constellation. After the linear equalizer (FFE) has reached a stable state or has been adjusted for a predetermined period of time, the adjustment of the linear equalizer is terminated and the training proceeds to the adjustment of the decision-feedback equalizer (DFE). After the decision-feedback equalizer (DFE) has been adjusted for a predetermined period of time or the decision-feedback equalizer (DFE) has reached a stable state, also the adjustment of the linear equalizer (FFE)

continued, whereby in this step both equalizers are being adjusted simultaneously.

After both equalizers have been adjusted for a predetermined period of time or a stable state of the system has been reached, the detector of the sparse symbol constellation is replaced by a detector corresponding to the symbol constellation used during normal data transmission. Inasmuch the points of the sparse symbol constella- tion coincide with the respective points of the symbol constellation used in the data transmission state, the switch-over between the detectors is straightforward. Subse- quently, receiver (RX) operates as if the symbol constellation of the data transmis- sion state is adopted in the system, while transmitter (TX) kind of"incidentally" continues to send data corresponding only to certain points of the symbol constella- tion. Hereupon, the switch-over of transmitter (TX) into the full symbol constellation of the data transmission state causes no transient effects in the receiver (RX).

If the system happens to include a Tomlinson-Harashima precoder (TML), the tap gain coefficients of the decision-feedback equalizer (DFE) are transmitted after the training period (or at the end thereof) via an auxiliary channel of the upstream transmission direction to the transmitter. The transmission of tap gain coefficient can be accomplished using either the sparse symbol constellation or the normal symbol constellation of the data transmission state.

Reference: [Lee & Messerschmitt] E. A. Lee and D. G. Messerschmitt, Digital Communication, Kluwer Academic Publishers, 1994.