Non intrusive channel impairment analyser.

The present invention relates to a non-intrusive channel-impairment analyser for measuring various quasi-static impairments in band-limited data channels without interruption of the channel traffic. More particularly, the present invention relates to such a channel-impairment analyser provided with means for modelling the channel under measurement.

Part of the subject matter disclosed herein also forms the subject of our co-pending European Patent Application No 87300040.0 from which the present Application is divided.

The use of band-limited voice channels for the transmission of high-speed data signals has been made possible in recent years by the development of a range of modems which incorporate data receivers that compensate for the major impairing effects encountered on such channels. These high-speed modems generally make use of two-dimensional modulation techniques whereby groups of binary data for transmission are encoded into two-dimensional symbols (made up of "in-phase" and "quadrature" components) which are then transmitted by amplitude modulating two quadrature carrier signals. The closed set of two-­dimensional data symbols characterises the type of modulation scheme; typical examples are phase shift keying (PSK) and quadrature amplitude modulation (QAM). The two-dimensional symbols can conveniently be represented on a two-­dimensional diagram (known in the art as a "constellation" diagram) with the in-­phase and quadrature components of each symbol being measured off along respective orthogonal axes. Figure 1 of the accompanying drawings illustrates a "16-QAM" constellation where each symbol represents four binary bits of a data signal being transmitted.

The constellation diagram shown in Figure 1 is, of course idealised. After the symbols have been transmitted over a voice-frequency data channel, the constellation diagram of the received symbols shows various distortions due to the effect of a range of quasi-static (or basically steady-state) and transient channel impairments.

Of the quasi-static impairments, linear distortion caused by the band-­limiting filter action of the channel is potentially the most troublesome and manifests itself as interference between adjacent symbols in the received data signal. Such intersymbol interference (ISI) is normally minimised by means of an automatic adaptive equaliser incorporated within the data receiver. Without such equalisation, however, the effect of channel linear distortion would be seen on a constellation diagram as a "cloud of uncertainty" clusterered around each constellation point (a similar effect is produced by additive background noise which co-exists with the data signal in the same spectral band). It should also be noted that as well as linear distortions, the channel will generally introduce second and third order non-linear distortions.

Two further significant quasi-static impairments are frequency offset and phase jitter which manifest themselves as a time-dependent variation in the reference phase of the data signal carrier wave. To minimise these effects, the data receiver normally includes a decision-directed phase-locked loop (PLL) which tracks variations in the reference phase of the data signal carrier wave. Without such a tracking device, the effects of frequency offset would be seen as a slow rotation of the constellation points about their origin, whilst the effect of phase jitter would be seen as an angular oscillation of the constellation points.

Another quasi-static impairment met in band-limited voice channels is amplitude jitter which manifests itself as a time-dependent variation in the level of the received data signal. It is possible to compensate for the effect of amplitude jitter by including within the data receiver a decision-directed gain-­control-loop which tracks variations in the level of the received signal. Without such a control mechanism, however, the effect of amplitude jitter would be seen on the data constellation diagram as a radial oscillation of the constellation points.

A known method of measuring channel impairments is to apply tones of predetermined frequency and amplitude to the channel and then observe the signal received at the far end. This method has the considerable disadvantage of requiring an interruption in the normal channel traffic.

Another known method of measuring channel impairments is described in US Patent No 4,381,546 (assignee Paradyne Corporation). This method concerns the obtaining of quantitative data on a channel over which data is transmitted by quadrature amplitude modulation. The method comprises the steps of: <ul id="ul0001" list-style="none"><li>a) producing sampled eye diagram point information wherein each received point is defined in a coordinate system in which a first axis is the in-phase channel axis and a second axis is the quadrature channel axis;</li><li>b) rotating said received points by an operand determined by the ideal value of the received point information so that each rotated point has its nominally maximum component on a new first axis and a second nominal component on a new second axis; and,</li><li>c) determining the characteristics of the communications channel from variances and means of the components.</li></ul>

It is an object of the present invention to provide a non-intrusive channel-­impairment analyser for measuring channel impairments by modelling the channel under measurement.

According to the present invention, there is provided a non-intrusive channel-impairment analyser for measuring at least one quasi-static impairment characteristic of a band-limited data communications channel, the analyser comprising: a data receiver section for receiving over said channel, data modulated onto quadrature phases of a carrier signal, the receiver section being arranged to process the received signal in two quadrature forward processing paths and including a data recovery circuit for effecting a decision as to the identity of the original data on the basis of the outputs from said processing paths, and decision-­directed compensation means disposed in said paths and arranged to compensate for channel-impairment effects on the received signal, and a measurement section responsive to signals generated in the receiver section during the receipt of random data, to derive a measurement of at least one said channel impairment, <b>characterised in that</b> said measurement section includes a real-time channel model comprising: an adaptive transversal model filter fed with the output of the data recovery circuit; demodulating means for demodulating an unequalised version of the signal received over the channel; comparison means for providing an error signal indicative of the difference between the output of the model filter and the output of said demodulating means; update means arranged to receive said error signal from the comparison means and to update the tap coefficients of the model filter such as to minimise said difference between the outputs of the model filter and the demodulating means; and Fourier transform means for deriving the channel linear characteristics from a Fourier transform of the tap coefficients of the model filter.

In deriving the model error signal, the comparison means of the channel model preferably includes third-order non-linearity compensation means arranged to modify said error signal by an amount dependent on the amplitude of the model output and a third-order non-linearity coefficient supplied to the non-linearity compensation means, whereby to reduce the effect on said error signal of third-order non-­linear distortion introduced by the channel.

Where third-order non-linearity compensation means are provided in the analyser, either in the model or elsewhere, then the analyser advantageously further comprises co­efficient determining means for adaptively determining said third-order non-linearity coefficient, the coefficient determining means being responsive to a.c. amplitude-error components of said error signal to update a previously derived value of said coefficient and this coefficient then being used by said third-order non-linearity compensation means.

A channel-impairment analyser embodying the present invention will now be particularly described, by way of non-­limiting example, with reference to the accompanying dia­grammatic drawings, in which: <ul id="ul0002" list-style="none"><li>Figure 1 is a 16-QAM constellation diagram;</li><li>Figure 2 is a block diagram of the channel-impairment analyser showing the main component blocks of a data-re­ceiver section and a measurement section of the analyser;</li><li>Figure 3 is a functional block diagram of a negative phase rotator used in the data-receiver section of the analyser to demodulate the received channel signal;</li><li>Figure 4 is a functional block diagram of a complex-­valued equaliser forming part of the analyser data-receiver section;</li><li>Figure 5 shows the structure of a loop filter used in a digital phase-locked loop of the data-receiver section;</li><li>Figure 6 is a functional block diagram of jitter and S/N ratio measurement blocks of the measurement section of the analyser;</li><li>Figure 7 shows the general structure of an adaptive line enhancer used in the jitter measurement blocks of Figure 6;</li><li>Figures 8a, 8b, 8c are in-phase/quad-phase component graphs illustrating the principle of operation of a second-­order non-linearity measurement block of the analyser meas­urement section;</li><li>Figure 9 is a functional block diagram of the second-­order non-linearity measurement section;</li><li>Figure 10 is a functional block diagram of a third-­order non-linearity measurement block of the analyser measurement section and illustrates the use of the output of that block to effect third-order non-linearity compensation in the determination of an amplitude error signal;</li><li>Figure 11 is an in-phase/quad-phase graph illustrating the significance of various error signals derived in the analyser;</li><li>Figure 12 is a functional block diagram of a channel model block of the analyser measurement section;</li><li>Figure 13 is a functional block diagram of an error generator used in the Figure 12 channel model;</li><li>Figure 14 is a functional block diagram of a third-­order non-linearity measurement block of the Figure 12 channel model; and</li><li>Figure 15 is a block diagram of a sampled-data proces­sor suitable for implementing the sampled data processing functions of the channel-impairment analyser.</li></ul>

The non-intrusive channel-impairment analyser is shown in Figure 2 in block diagram form. As can be seen, the analyser is made-up of two sections, namely a data receiver section (upper half of Figure 2 above dashed line) and a measurement section (lower half of Figure 2).

THE RECEIVER SECTION

The data receiver section of the analyser is based on the receiver type disclosed in US Patent Specification No. 3,878,463 (Falconer et al) and, more particularly, corres­ponds closely to the implementation of that receiver type described by Falconer in Bell Systems Technical Journal Vol55, No.3, March 1976 at page 323. This latter imple­mentation has, however, been modified in the Figure 2 receiver section to incorporate a second order loop filter in the PLL used for phase correction, this modification being in accordance with the proposal by Levy and Poinas in a paper entitled "Adaptive Phase Correctors for Data Trans­mission Receivers".

As a detailed explanation of the functioning of the receiver section can be found in the source documents refer­red to above, only an outline of the receiver operation will be given herein to assist in the understanding of the nature of the signals passed to the measurement section of the analyser.

The present analyser is intended for use with voice-­frequency channels carrying two-dimensional symbols modu­lated onto orthogonal phases (sinω_ct and cosω_ct) of a common carrier of frequency ω_c at a rate I/T where T is the baud (symbols per second) interval.

During each baud interval the data to be transmitted can be represented by the numbers I and Q, being the inphase and quadrature components respectively of the corresponding symbol. The numbers I and Q can be considered as the components of an analytic signal R where R = I + jQ

The process of modulating the components I and Q into orthogonal phases of a common carrier may then be viewed as a multiplication of the signal R by the complex frequency e^jωc^t to form a new complex signal S and then taking the real part as the real signal to be transmitted: S = R e^jωc^t = (I + jQ) e^jωc^t = (Isin^ω_ct-Qcos^ω_ct)+j(Isin^ω_ct+Qcos^ω_ct)

In the receiver section of the analyser, the received signal is first passed to analog signal conditioning circuitry 10 (including gain auto-ranging) before being sampled in sampler 11 at an even integer multiple of the symbol rate under the control of a symbol-clock recovery circuit (not shown). Subsequent signal processing is effected digitally either by circuitry dedicated to each processing function to be performed or by a suitably programmed, general purpose sampled data processor.

To recover the components I and Q in a data receiver supplied with the transmitted signal (Isin^ω_ct - Qcos^ω_ct), an analog to the imaginary part of the complex signal S is first derived by passing the received signal through a phase splitter made up of a pair of transversal filters 12, 13 operating at the sampling frequency. These filters are matched to the transmitted signal but differ in phase shift by 90^o. The resulting outputs from the filters 12,13 correspond to: Filter 12: Isin^ω_ct-Qcos^ω_ct Filter 13: Isin^ω_ct+Qcos^ω_ct that is, to the real and imaginary components of S. The two filter outputs taken together thus represent R.e^jωc^t.

The components I and Q can now be recovered by multiplying the quantity R.e^jωc^t by e^⁻jωc^t leaving R or, in other words (I + jQ). Multiplication by e^⁻jωc^t corresponds, of course, to a phase rotation by -ϑ^owhere ϑ = ω_ct, Figure 3 illustrates, in diagrammatic form, the implementation of a negative phase rotator for rotating a complex quantity (A + jB) by -ϑ^o This rotator comprises a sin/cos generator 2 for generating the quantity sinϑ and cosϑ correspond­ing to an input ϑ; the generator 2 may, for example, be in the form of a look-up table (if only predetermined selec­ted values of ϑ are to be used) or means for generating sinϑ and cosϑ by a power series expansion. The Figure 3 rotator also includes multipliers 3 and adders 4 and it can be readily seen that the outputs of the rotator corres­pond to the real and imaginary parts of the multiplication (A + jB) e^-jϑ.

In the present receiver section, the output of the phase splitter (filters 12,13) is passed to a phase rotator 15 of the Figure 3 form, which is arranged to effect a phase rotation corresponding to (-ω_dt) where ω_d is nomin­ally equal to ω_c. The phase rotator 15 is, of course, digital in form and operates at twice the baud rate.

In the absence of channel impairments, (in particular, frequency offsets and phase jitter), the output of the phase rotator 15 would be the signal R from which it will be appreciated that the rotator 15 serves as a demodulator.

In practice, the existence of channel impairment means that further processing of the complex-valued demodulator output signal X (in-phase and quadrature components x_i and x_q respectively) is required before the signal R can be reliably recovered. Following the demodulator 15 there are therefore provided two channel compensation stages, namely a channel equaliser 17 and an excess carrier phase compensator in the form of a phase rotator 16. As will be more fully explained below, both these compensation stages are decision-directed; that is, the compensation provided thereby is based on a decision as to the values of the transmitted data; this decision is taken by a slicer 20 which immediately follows the rotator 16.

The channel equaliser 17 takes the form of a complex-­valued transversal filter, operated at twice baud, whose tap coefficients (in-phase and quadrature) are updated in such a way as to minimise a measure of the complex valued intersymbol interference introduced by the channel. Figure 4 diagrammatically depicts the processing structure of a complex-valued transversal filter and comprises four con­volution blocks 18 and two summers 19 for generating the following output components (in vector notation) making up the complex-valued output signal Y. In-phase output y_i = c_i^Tm_i-c_q^Tm_q Quad-phase output y_q = c_i^Tm_q+c_q^Tm_i where: c_i^T and c_q^T are the transposed column vec­tors of the in-phase and quadrature tap-gain coefficients respectively, m_i is the column vector of in-phase samples taken at taps along the in-phase delay line, and m_q is the column vector of quadrature-phase samples taken at taps along the quadrature phase delay line. The updating of the tap coefficients will be dealt with hereinafter.

The output Y of the adaptive channel equaliser is fed to the phase rotator which forms the excess carrier phase compensator 16. This rotator 16 forms part of a decision-­controlled phase-locked loop intended to remove excess carrier phase rotation of R caused by frequency offset and phase jitter in the channel. The updating of the value of the phase rotation -ϑ_o effected by the phase rotator 16 is explained more fully below. The output of the rotator 16 is a complex-valued signal Z with in-phase and quadrature components of z_i and z_q respectively.

In spite of the compensation effected by the equaliser 17 and rotator 16, the complex-valued signal Z remains corrupted by residual uncompensated channel impairment effects such as background noise. The final stage of the data retrieval processing is, therefore, a slicing operation performed by the previously-mentioned slicer 20. The slicer 20 operates as a multi-level threshold detector for the in-­phase and quadrature channels to provide a best estimate D of the transmitted two-dimensional data symbol; the in-phase and quadrature components, d_i and d_q respectively of the signal D are thus best estimates of the components I and Q of the transmitted signal. The form and operation of the slicer 20 are well understood by persons skilled in the art and will not be further described herein.

The updating of the channel compensation stages formed by the equaliser 17 and rotator 16 is effected in dependence on error signals derived by comparing the input and output signals Z and D of the slicer 20.

More particularly, to update the tap coefficients of the equaliser 17, a complex-valued vector error signal E having in-phase and quadrature components e_i and e_q respectively, is first derived in error generator block 21 in accordance with the following expressions: e_i = z_i - d_i e_q = z_q - d_q

Figure 11 shows the inter-relation of the signals Z, D and E. Before this vector error signal E can be input into a tap update algorithm, it must be fed through a positive phase rotator 22 to reverse the effects of the excess carrier phase compensator 16 (the equaliser 17 being upstream of the compensator 16). The rotator 22 is similar in form to that shown in Figure 3 but with the negative input to the lower summer 4 made positive and the corresponding input to the upper summer 4 made negative; this change is necessary to produce a positive rotation of O- rather than the negative rotation effected by the Figure 3 circuit. The magnitude of rotation ϑ_o effected by the phase rotator 24 is thus equal and opposite to that effected by the compensator 16. The resultant error signal F (having inphase and quadrature components of f_i and f_q respectively) is then used in a tap-update algorithm (block 23) of known form such as that described in the above-mentioned US Patent Specification No 3,878,468; according to the algorithm described in this latter document, the tap coefficients c_i,c_q are adjusted in conformity with the following expressions: c_i(n+1) = c_i(n) - β (f_im_i+f_qm_q) c_q(n+1) = c_q(n) - β (F_im_q-f_qm_i) where n, (n+1) refer to successive baud intervals and β is a constant.

For the adaptive equalisation process to function properly and the equalizer tap coefficients to converge to their optimum values, the transmitted data must be random (this will generally be the case where a data scrambler is provided at the data transmitter). Once equaliser conver­gence has been attained, the adaptive process allows the equaliser 11 to track slow variations in the channel characteristics; one result of this is that the equaliser serves as a narrow-band, decision-directed, gain control system. However, due to the range of possible interfering effects present in the error signal, the tap coefficients will exhibit some degree of uncertainty which results in a component of residual intersymbol interference in the received signal. Consequently, a compromise must be made in the selection of the update constant β which determines both the tracking capability and the tap coefficient stability of the equaliser.

The updating of the phase-angle value ϑ_o fed to the excess carrier phase compensator 16 involves the deri­vation of a phase error signal ϑ_e by subtracting the phase angle of the slicer output from the phase angle of the slicer input (see Figure 11). The slicer output phase angle corresponds to:<maths id="math0001" num=""><img file="EP0366159A1_D0001.tif" /></maths> This quantity can be held in a look-up table since there are only a limited number of slicer output combinations. On the other hand, the slicer-input phase angle:<maths id="math0002" num=""><img file="EP0366159A1_D0002.tif" /></maths> must be computed by a power series expansion as the in-­phase and quadrature components of the slicer input may be in any relation. The derivation of the phase error signal ϑ_e is effected in the error generator block 21 of Figure 2.

The phase error signal signal ϑ_e is passed to a loop filter 24 of the digital phase-locked loop composed of the rotator 15, error generator 21 and filter 24. The loop filter 24 is shown in greater detail in Figure 5 and, as can be seen, the filter 24 is composed of summers 25, multi­ pliers 26, and baud-interval (T) delays 27, arranged as a second order filter for providing a narrow-band phase output ϑ_e to control the phase rotator 16. More partic­ularly, the output ϑ_o is updated in accordance with the following difference equations:<maths id="math0003" num=""><img file="EP0366159A1_D0003.tif" /></maths> and ϑ̇_o(n) = ϑ_o(n-1)+α₂φ_e(n) where (n-1) ,n, and (n+1) refer to successive baud interval values of the corresponding quantities, and α₁ and α₂ are constants selected to maximise the training and tracking performance of the phase-locked loop.

The integral control introduced by the second order loop dynamics is present specifically to compensate for frequency offset which appears as a ramp in phase; the quantity ϑ̇_o is, indeed, a measure of frequency off­set.

THE MEASUREMENT SECTION

As shown in Figure 2, the measurement section of the channel-impairment analyser comprises: - a phase-jitter and frequency offset measurement block 30: - an amplitude jitter measurement block 31; - a signal/noise ratio measurement block 32; - a 2nd order non-linearity measurement block 33; - a 3rd order non-linearity measurement block 34; and - a channel model block 35. These measurement blocks will be described in turn below.

Phase Jitter and Frequency Offset

- The phase jitter and frequency offset measurement block 30 is shown in more detail in Figure 6. The first operation carried out in block 30 is to form a wideband phase error signal by combining in summer 40, the phase error signal ϑ_e from the error generator 21 and the narrow-band output signal ϑ_o from the PLL loop filter 24. This latter signal re-introduces into the phase error signal, inform­ation on low-frequency phase error removed by the action of the rotator 16 thereby enabling wideband phase-jitter measurements to be made which would not be the case if the signal ϑ_e had been used alone.

In fact, in the present analyser, phase jitter measure­ments are carried out in two frequency bands, namely between 4 and 20Hz and between 20 and 300 Hz. To this end, the wideband phase error signal is first passed through a 2Hz narrowband digital phase-locked loop 41 to provide the 4Hz lower measurement limit, and then through a transversal low-­pass filter 42 providing the 300Hz upper measurement limit. Thereafter, the 4-300Hz band-limited signal is fed to two separate measurement channels that correspond to respective ones of the measurement bands. The 20-300Hz channel comprises a 10Hz DPLL providing a 20Hz high pass filter 43, and a peak-to-peak detector 44. The 4-20Hz channel comprises a 20Hz transversal low-pass filter 45 followed by a peak-to-peak detector 46.

The DPLL is of a form similar to that shown in Figure 5 for the loop filter 24 but with the constants ₋₁ and ₋₂ chosen to give the desired frequency-limiting charac­teristic rather than to give a particular tracking and training performance. The DPLL 41 is also used to derive a frequency offset signal in the manner previously described with reference to Figure 5 (the desired signal being that corresponding to the signal ϑ_o(n) of Figure 5). The frequency offset signal is passed to an averager 47 before being output.

As can be seen from Figure 6, the 20-300Hz phase jitter channel is provided with an adaptive line enhancer 48 the purpose of which is to remove background noise and other non-deterministic components in the phase error signal in this channel (the provision of a similar enhancer in the 4-20Hz channel has been found to be unnecessary since the noise power in this narrower frequency band is much less). The problem of noise swamping the deterministic phase error components is particularly severe in the case of QAM constellations which have a set of inner data points as well as a set of outer data points (see, for example, the 16-QAM contellation of Figure 1) since the same noise power produces a much larger phase error for the inner data points than the outer ones.

The form of the adaptive line enhancer is illustrated in Figure 7. As can be seen, the enhancer comprises a delay (ΔT) 50 fed with the noise-corrupted jitter signal J_IN(n), followed by a single-valued adaptive transversal filter 51 whose taps are spaced at the baud interval T. The input signal J_IN(n) is compared in a comparator 52 with the output signal J_OUT(n) from the filter 51 and the resultant error signal E_J(n, is supplied to a tap co­efficient update block 53. The block 53 implements a tap update algorithm of the same general form as used for the adaptive equaliser 17 but simplified for real components only.

The combined effect of the delay ΔT and the filter phase shift is such that the deterministic components in the input and output signals of the line enhancer are phase coherent and cancel at the comparator 52. The delay ΔT is chosen to ensure that there is no correlation between likely interfering signals in the input and output of the enhancer 48 so that these will not cancel at the comparator 52. Thus, at convergence, the error signal E_J(n) should consist only of any interfering signal present at the input to the line enhancer 48 thereby giving the line enhancer a spectrum with narrow pass bands centered at the frequencies of the deterministic components. In practice, however, a small component of the deterministic signals appears in the error signal due to the filter output being slightly attenuated. It can in fact be shown that as the number of taps is increased, there is an increase both in the accuracy of the enhanced signal and in the frequency resolution of the filter.

Further description of this form of adaptive line enhancer may be found in an article entitled "Adaptive Noise Cancelling : Principles and Applications" (Widrow et al), page 1692, Vol.63, No.12, Proceedings of the IEEE, December 1975.

As previously noted, the sampled data processing effected in the analyser may be carried out by circuitry dedicated to each particular function indicated in the accompanying Figures or by a general purpose sampled data processor. In this latter case, time constraints will normally make it difficult to update the tap coefficients of the line enhancer 48 during every baud interval and updating will generally therefore be done on a multiplexed basis with other processing tasks which do not absolutely require to be effected every baud interval. It has, however, been found that the line enhancer 48 is susceptible to producing mis­leading results if this multiplexed tap updating is carried out at fixed, regular intervals; the reason for this is that the interval chosen may be related to the frequency of a deterministic jitter component with the result that updating repeatedly occurs at the same phase of that component and the latter is therefore effectively ignored. To overcome this difficulty, the number of baud intervals between updating of the line enhancer 48 is randomised in any suitable manner.

Amplitude Jitter

- The amplitude jitter measurement block 31 is shown in more detail in Figure 6. The input to block 31 is a wideband amplitude error signal A_e (see Figure 11) formed in the error generator 21 by subtracting the amplitude of the slicer output from that of the slicer input:<maths id="math0004" num=""><img file="EP0366159A1_D0004.tif" /></maths> The term ( d_i² + d_q²)^½ can be obtained using a look-up table as there are only a limited number of possible outputs from the slicer 20. The term ( z_i² + z_q² )^½ is most readily determined in the form:<maths id="math0005" num=""><img file="EP0366159A1_D0005.tif" /></maths> as the quantity z_q/z_i will have been previously calculated during derivation of the phase error signal O_e (see before).

In fact, the signal A_e will not include very low frequency components as these will generally have been tracked out by preceding stages of the data receiver.

As will be described hereinafter, the amplitude error signal is also subject to compensation for third order non-­linear distortion (NLD) prior to being passed to the ampli­tude jitter measurement block, this compensation being effected in the error generator 31 as a result of feedback from the third-order non-linearity measurement block 34.

The processing of the wideband amplitude error in block 31 is similar to the processing of the phase error in block 30. More particularly, the amplitude error, after passage through a 300 Hz transversal, low-pass filter 59, is processed in two channels, one for frequencies in the range 4-20Hz and the other for frequencies in the range 20300Hz. The 4-20Hz channel comprises a 2Hz DPLL 60, a 20Hz, transversal, low-pass filter 61, and a peakto-peak detector 62. The 20-300Hz comprises a 10Hz DPLL 63, an adaptive line enhancer 64 similar to the line enhancer shown in Figure 7, and a peak-to-peak detector 65.

Signal-to-Noise Ratio

The S/N ratio measurement block 32 is shown in more detail in Figure 6. The computation of S/N ratio is based on the approximation: S/N = [(Phase error)² + K(amp error)²]K_scale

In order to exclude jitter modulation components from inclusions in the noise, only phase and amplitude error components above 300Hz are used to calculate the S/N ratio (it being expected that modulation components will generally be below this frequency). To generate the required phase error signals, the output of the 300Hz low-pass filter 42 is subtracted from the filter input in block 66; similarly the required amplitude error signal is generated by subtracting the output of the 300Hz low-pass filter 59 from the input to the filter in block 67. The outputs of blocks 66 and 67 are passed to a processing block 68 for computation of the S/N ratio in accordance with the expression given above. To compensate for the fact that the bottom 300Hz have been removed from the error signals, a scaling factor is intro­duced.

As mentioned above, for a QAM constellation including both inner and outer sets of points, the effect of noise is exaggerated for the inner set of points. For such constel­lations, therefore, a further scaling factor is introduced to correct for this effect.

2nd Order Non-Linearity

- While linear distortions introduced by the channel will be largely compensated for by the equaliser 17, this is not the case for second and third order non-linear distortions.

With regard to second-order non-linear distortions, the Applicants have observed that such non-linearities can be revealed and measured by remodulating the vector error components e_i,e_q (see Figure 2) of the outer corner points of a QAM constellation (see points ringed by dashed circles in Figure 8a) up to the frequency at which 2nd order non-linearities may have been introduced. Assuming a random distribution of the original data points, the remodu­lated error components will either: - produce a random distribution of points having no particular bias (that is, if the remodulated error points are plotted on an in-phase component/quad-phase component graph, the resultant cluster of points will have its centroid at the origin as indicated in Figure 8b); or - produce a biased distribution of points with the centroid of the points cluster offset from the origin of an in-phase/quad-phase component graph (see Figure 8c). This latter condition is indicative of the presence of second-order non-linearities with the magnitude of the offset being a measure of the non-linearities.

Owing to the fact that the channel may introduce a frequency offset either before or after the introduction of second-order non-linearities, the frequency to which the vector error components must be remodulated will not necessarily be the frequency of the signal received at the analyser. This problem can be overcome by remodulating the error components to the received-signal frequency (carrier nominal plus detected offset) and then slowly sweeping the remodulation frequency through a few Hertz either side of the received-signal frequency while monitoring the position of the points cluster centroid. If any second-order non-­linearities are present, then the points cluster centroid will exhibit an offset which peaks at a particular value of the remodulation frequency; the centroid offset at this frequency can be considered as an indication of the magni­tude of non-linearity present.

Figure 9 shows, in functional block diagram form, a processing arrangement for implementing the foregoing method of detecting second-order non-linearities. The vector error components e_i,e_q of each outer corner point that occurs (as detected by the slicer 20), are fed first to a positive phase rotator 70 for remodulation to the nominal carrier frequency ω_d and then to a positive phase rotat­or 71 for further modulation in correspondence to the detec­ted frequency offset (as embodied in the signal ϑ_o). The in-phase and quad-phase components output from the rota­tor 71 are thus at the frequency received by the analyser. These components are then passed to two separate channels 72,73 respectively arranged to increase and decrease the modulation frequency in ¼Hz steps from 0 to 7Hz and to search for any resultant centroid offset.

The channel 72 comprises a positive phase rotator 74 controlled by the output of an integrator 75 that is fed with a phase angle value φ_INC. For constant φ_INC, the phase angle supplied to the rotator 74 will increase simulating a frequency determined by the magnitude of φ_INC. By changing the magnitude of φ_INC at intervals, the frequency produced can be swept between 0 and 7Hz in the desired ¼Hz steps. The in-phase and quad-phase outputs of the rotator 74 are passed to respective averaging filters 76,77 and then to respective squaring circuits 78,79 before being summed in summer 80. The output of the summer 80 feeds a low-pass output filter 81.

The channel 73 is similar to the channel 72 except that a negative phase rotator 82 is provided instead of the posi­tive phase rotator 74; as a result, progressive increases in φ_INC result in the frequency change introduced in channel 73 being in a negative sense. The remaining compo­nents of channel 73 are the same as for channel 72 and have, therefore, been similarly referenced in Figure 9.

If a second-order non-linearity has been introduced by the channel into the originally transmitted signal, then as φ_INC is increased to increase the frequency change introduced by the rotator 74,82, the output of one of the channels 80 or 81 will increase to a maximum as the fre­quency in that channel matches the frequency at which the distortion was introduced. This maximum is taken as a measure of the magnitude of the second-order non-linearity present in the channel.

3rd Order Non-Linearity

- The presence of third-order non-linearities is most noticeable in its effect on the outer corner points of a QAM constellation, these points being significantly radially expanded or compressed in the presence of such non-linearities. To compensate for this effect, the previously stated expression for calculating the wideband amplitude error A_e is modified to:<maths id="math0006" num=""><img file="EP0366159A1_D0006.tif" /></maths> where C₃ is a third-order non-linearity coefficient. The functional blocks required to derive A_e in accor­dance with the above expression are illustrated in the upper half of Figure 10 and comprise processor blocks 90,91,92, summers 93,94, multiplier 95, and divider 96 (the purpose of the latter being to normalise the derived value of A_e by division by the amplitude value of the sliced point).

The derivation of the third-order non-linearity co­efficient C₃ is effected by the functional blocks illus­trated in the lower half of Figure 10. The normalised amplitude error A_e is first passed to a digital phase-­locked loop made up of summer 97 and an integrating loop filter 98 with update constant K₁, to remove any dc component from the amplitude error (this loop is updated every baud interval). The average output from the summer 97 is thus zero. If an outer, corner constellation point has been sliced, the output from the summer 97 is used to update the value of the third-order non linearity coefficient C₃ by means of an integrator 99 constituted by a first-­order DPLL with an update constant K₂ provided by an input multiplier 100. The constants K₁ and K₂ are small.

If no third-order non-linearity is present in the received signal then C₃ will converge to zero. However, if compressive third-order non-linearities are present, C₃ will converge to a negative value, whereas if expan­sive third-order, non-linearities are present, C₃ will converge to a positive value.

Channel Model

- The channel model block 35 is provided to facilitate the measurement of channel linear distortion.

The general arrangement of the channel model block 35 is shown in Figure 12. The model is built around an adapt­ive complex-valued model filter 110 which is fed with the output signal D (components d_i,d_q) from the data-­receiver slicer 20. The taps of the filter 110 are adjusted to minimise the error between the output of the model filter 110 and the output X (components x_i,x_q) of the data-­receiver demodulator 15 after rotation of this latter output to compensate for excess carrier phase ϑ_o (this com­pensation is effected by a negative rotator 111). Once convergence of the model filter 110 has been attained, the filter models the channel linear distortion and will track slow variations in the channel characteristics. By per­forming a Fast Fourier Transform on the tap coefficients of the model filter 110, the channel's frequency characteristic can then be derived (this is represented by block 112 in Figure 12).

It will, of course, be appreciated that the signals X, D and ϑ_o required by the model must be delay-matched before being supplied to the model.

The form of the model filter 110 is substantially the same as that illustrated in Figure 4 for the equaliser 17. To avoid aliasing, the model processing is carried out at twice baud rate. In practice, the complex-valued input signal supplied to the model is obtained by sampling the output of the slicer 20 at T-spaced intervals and interleav­ing these samples with zero samples. This has the advan­tage that alternate outputs are associated with odd and even taps, so for the purposes of tap updating the filter 110 can be considered as two filters each with half the number of taps. The updating process itself (block 113 in Figure 12) is effected in accordance with the same algorithm as is used for the equaliser 17.

The generation of the model vector-error signal E_M (components of e_Mi,e_Mq) supplied to the tap coeffi­cient update block 113 is effected in error generator 114 and includes compensation for third-order non-linearity as represented by a coefficient C_3M derived in block 115.

Designating the output of the rotator 111 by the complex variable G (components of g_i and g_q) and the output of the model filter 110 by the complex variable H (components of h_i and h_q), the model error E_M is derived in accordance with the following expression: Model Error = Received point - [Model o/p+3rd NLD comp.] (e_Mi+je_Mq) = (g_i+jg_q) - [(h_i+jh_q) + C_3nr³(cosφ+jsinφ)] where r and φ are respectively the amplitude and phase of the model filter output, Since:<maths id="math0007" num=""><img file="EP0366159A1_D0007.tif" /></maths> the preceding expression can be written as: (e_Mi+je_Mq) = (g_i+jg_q) - [(h_i+j)h_q + C_3M(h_i²+h_q²) (h_i+jh_q)] This expression is, of course, not of the same form as that given earlier in relation to 3rd order NLD compensation of the amplitude error A_e in block 21; this difference is due to the wholly different nature of the error signals concerned as may be appreciated by reference to Figure 11 where the present vector error signal E_M is akin to the vector E rather than the scaler A_e.

Figure 13 shows in detail the main processing blocks of the error generator 114. These blocks comprise a block 120 for deriving the quantity r², multipliers 121,122,123 and summers 124,125,126 and 127.

A similar compensation for third-order non-linearities may also be efffected in deriving the vector error E in the receiver section of the analyser.

The model third-order non-linearity coefficient C_3M is derived in block 115 the same way as the coefficient C₃ is derived in block 34. Thus, referring to Figure 14, the block 115 can be seen to include a processing portion 132 corresponding to the part of Figure 10 that illustrates the arrangement for deriving C₃ from the wideband amplitude error A_e. The processing portion 132, like the corresponding portion of Figure 10, requires for its input a scaler signal representing amplitude error rather than the vector error. Accordingly, the block 115 is provided with an arrangement for converting the model vector error E_M (components e_Mi,e_Mq) into a model amp­litude error A_Me. The principle behind the conversion operation can be understood by reference to Figure 11 in which the variables D,Z,E and A_e bear the same general inter-relationship as the variables H,G,E_M and A_Me of the model. If, for the sake of simplicity, the 3rd order non-linear distortion compensation term in the vector error E_M (analogous to E of Figure 11) is ignored, then, the model amplitude error A_Me (A_e, Figure 11) can be obtained by rotating the vector error clockwise by the phase angle of the vector H (D, Figure 11) and taking the real part: A_Me = R_e [ (e_Mi=je_Mq) (cosΨ - jsinΨ)] where Ψ is the phase angle of the vector M.

This rotation through Ψ is effected in Figure 14 by the negative rotator 131. However, since the calculation of cosΨ and sinΨ by a power series expansion involves signifi­cant processing time, in the present embodiment rather than computing exact values for cosψ and sinψ, an approximate angle value is derived by means of an eight-phase slicer 130 fed with the signal components h_i and h_q. The in-­phase and quadrature output components are then used directly in the rotator 131 without the need for a sin/cos generator. With this arrangement, there is a possibility of a rotation error of 22.5^o maximum.

The output of the C_3M derivation block, as well as being used in the model error generator 114, may also be made available externally.

As is illustrated in Figure 12, the model may also be provided with a block 116 for deriving a model second-order non-linearity measurement. This block operates in the same general manner as the block 33 and will, therefore, not be further described herein.

SAMPLED-DATA PROCESSOR

As previously mentioned, the digital signal processing effected in the data receiver and measurement sections of the analyser may be implemented either by circuitry dedi­cated to each processing function or by means of a real-time sampled-data processor.

Where a sampled-data processor is used, the internal architecture of the processor is preferably optimised for the range of tasks to be performed and a suitable architec­ture for the present case is illustrated in Figure 15. Since the principles of operation of such sampled-data processors and their detailed design are well understood by persons skilled in the relevant art, only a brief descrip­tion of the Figure 15 processor will be given below.

The Figure 15 processor basically, comprises two parts, namely a control unit and an execution unit. The control unit is based upon micro-programming techniques where the micro-program instruction words (micro-instructions) reside in memory 200 and an instruction sequencer 201 generates the memory address of the instruction word to be carried out. Each micro-instruction consists of the range of signals required to control both parts of the processor. The sequencer 201 includes a counter 202, stack 203 and microprogram counter 204.

Operationally, the processor may be described as a 'data-driven machine' since only the data transfer rate (input and/or output) with the analog parts of the analyser determines the micro-program repetition period. Synchro­nising to the data rate is accomplished by first-in first-­out (FIFO) memories 202, 203 which link a main data bus 205 of the signal processor with the analog interface. Each program cycle is carried out at the full speed of the processor (nominally 6 instructions per micro-second) unless the input FIFO 202 is empty or the output FIFO 203 is full. In these latter two cases the program 'waits' for the appropriate data availability condition being met before continuing.

The processor execution unit is a collection of elements whose architecture has been designed to carry out the sampled data processing activities of the analyser as efficiently as possible in real-time. Specifically, the execution unit has been optimised to carry out the function of digital convolution as efficiently as possible since this function represents the largest percentage requirement for real-time computation. For this reason, the execution unit is built around a high speed 16 x 16 multiplier-accumulator 210 arranged to be fed with filter tap coefficients and signal sample data simultaneously for fastest operation. The multiplier-accumulator 210 includes two input registers 213,214 respectively designated X and Y which feed a multi­plier 215. The multiplier 215 is followed by an adder 216 that enables the progressive accumulation of processing results in a register 217. To achieve the simultaneous supply of tap coefficients and signal sample data to the multiplier-accumulator, the execution-unit memory has been split into two segments, called the 'data' memory 211 and the 'coefficient' memory 212 as shown. The data memory 211 is connected directly to Y input register 214 of the multi­plier 210 and via bi-directional bus drivers 218, to the 15 main data bus 205, whilst the coefficient memory is connect­ed directly to the main data bus 205 and then to the X input register 213 via bi-directional bus drivers 219. In con­junction with the control capability of the sequencer 201, this arrangement allows a single execute-and-loop instruc­tion to be assembled for fastest operation of the digital convolution function. An additional small overhead is required to initialise the computation by setting the filter length in the sequencer 201 and initialising address coun­ters 219, 220 respectively associated with the data memory 211 and coefficient memory 212, (the initialisation data being held either in memory 212 or in a scratchpad memory 221).

The counters 219 and 220 have the capability to cycle on a preselected fixed depth of memory which minimises the overhead required in a data storage and access; both these latter activities are controlled by a single pointer which may be held either in the scratchpad 221 or in the memory 212. The cycle depth of each counter 219,220 is indivi­dually controlled via their respective mask registers 222, 223. Additionally, the least significant bit of counter 219 may be reversed which allows complex-valued data to be stored interleaved under the control of a single pointer and thereafter to be readily unscrambled for the purpose of complex digital convolution.

The majority of the filtering activities carried out in the processing are adaptive. However, the complexity of the tap update algorithms is such that, although each filter convolution must be carried out at the required rates to maintain receiver integrity, the update procedures for each adaptive filter are multiplexed in time with only one filter being updated each symbol period. For tap coefficient updating, the error signal is held in scratchpad 221 so that, within a loop structure for multiple tap updates, counter 220 addresses the coefficient to be updated and counter 219 addresses the data component in the update algorithm. With this computational organisation, the processor architecture allows a complex-valued tap coeffi­cient to be updated within a loop structure of eight instructions.

Another extensively used algorithm is the power series expansion of mathematical functions. For this algorithm, the polynomial coefficients are held in the coefficient memory 212 and the independent variable of the function is preloaded into one of the input registers of the multiplier 210, which allows the algorithm to be implemented using the process of repeated factoring in a loop structure requiring three instructions. Thus a fifth order polynomial may be executed in six cycles of three instructions plus a small overhead for initialisation purposes.

For products involving complex numbers, one variable is usually held in the scratchpad memory 221 which provides access directly to one input of the multiplier 210 so that the other variable may be addressed from either the data memory 211 or the coefficient memory 212 for access to the other free multiplier input. In addition, the scratchpad 221 is used to hold parameters such as data points and constant values which are common to the activities of several subroutines, and also to provide an area of working storage for other special subroutine activities such as autorange control and timing control.

An interface between the control unit and the execution unit is provided by means of registers 230,231. Register 230 provides the link 'execution-unit to control-unit' whilst register 231 provides the link 'control-unit to execution-unit'. Both the registers 230,231 may be used to determine sequencer branch instructions conditionally, on the basis of a comparison (effected in comparator 232) with the execution-unit data bus 205, and unconditionally, by value passed in micro-program via register 231 or by value passed from the execution-unit via register 230. In addition, the register 231 allows the sequencer counter 202 to be initialised for applications which require a fixed loop count or a branch to a fixed location whilst register 230 allows the sequencer counter 202 to be initialised dynamically under signal processing/execution control for applications which require a variable loop count.

From the foregoing, it can be seen that although the architecture of the Figure 15 processor has been optimised for fast digital convolution (which is the most time consum­ing of analyser activities that have to be carried out in real-time every data symbol interval), it incorporates features which allow the other signal processing activities to be carried out very efficiently with the minimum of non-­computational overhead. Indeed, the processor may be regarded as a general purpose device suitable for a range of sampled-data signal processing activities.