GB2314241A - Data symbol estimation - Google Patents

Abstract

A digital mobile radio communications system and, a neural network method for operation thereof includes a radial basis function (RBF) network 5 in which an estimated data symbol is output based on a parameter D MAN The Manhattan distance parameter D MAN is determined by the equation ```where N e is the dimensionality of the observation vector, X i is the i th entry in the observation vector, and C i is the i th entry in the state center vector. Computational cost is reduced, as no multiply operations are required , and digital VLSI implementation becomes possible. Application is to channel equalisation, diversity combining or joint detection of spatial division multiple access (SDMA) datastreams, in the context of GSM cellular telephone systems.

Description

DIGITAL COMMUNICATIONS SYSTEM The present invention relates to a digital communications system, and to methods of operation for such a system. More particularly it relates to the recovery of data from signals transmitted on digital communications channels.

In the context of digital mobile radio systems, for example, there are three main areas where signal processing for recovery of data is required: channel equalisation in which a transmitted data symbol is estimated from successive samples of a noisy received signal; diversity combining in which the transmitted data symbol is estimated from simultaneus samples of the noisy received signal from an array of receivers; and joint detection, in which the transmitted data symbols of more than one source are simultaneously estimated from the noisy samples of the received signals from an equaliser of diversity combiner.

The first two of these areas provide techniques for improving the quality of a radio link and reducing outage due to adverse propagation effects, whereas the third is a technique for increasing the capacity of cellular radio systems by allowing more than one user to share a physical channel.

Conventionally signal processing to recover data has been carried out using linear filtering. More recently the use of non-linear techniques and particularly techniques based upon the use of neural networks has been propsed.

Recognising the finite state nature of digital modulation schemes, the problem of symbol recovery has been reformulated as a classification problem, making possible the use of neural networks. Several types of non-linear architecture have been proposed including multi-layer perceptron networks, Volterra series and Radial Basis Function (RBF) networks. Of these, RBF networks are potentially particularly promising because the structure of this network is closely matched to the problem. However, to handle the data rates typical of digital mobile radio, which may of several hundred kbit/s, a hardware implementation is required, and hitherto the complexity of RBF networks has been such as to make a hardware implementation prohibitively expensive.

According to a first aspect of the present invention, there is provided a method of operating a digital communications system including applying a noisy signal to a radial basis function (RBF) network comprising a plurality of nodes corresponding to different symbol values and outputting from the network an estimated data symbol, characterised in that the nodes of the network compute a parameter D8N, where

and in that the network selects the value of the estimated data symbol according to the different respective values of DAN for the different nodes.

The present invention provides a new form of RBF network which for the first time makes practical the implementation of such a network in communications systems.

The present inventors have a developed a new basis for the calculation of the radial function, using the Manhattan distance DMAN, which greatly simplifies the structure of the network while producing little loss in performance. As will be further explained below, the approach of the present invention eliminates the need for multiply operations in the network and reduces the computational cost to 2Ne additions per node, where Ne is the number of dimensions of the sample array.

The sample array input to the network may comprise successive samples of a noisy signal from a single receiver, or alternatively may comprise simultaneous samples from an array of receivers.

The method may included joint detection of transmitted data symbols from more than one source.

According to a second aspect of the present invention there is provided a receiver for a digital communications system including a radial basis function (RBF) network comprising a plurality of nodes corresponding to different symbol values and arranged to output from the network an estimated data symbol, characterised in that the nodes of the network are arranged to compute a parameter DN, ' where

and in that the network selects the value of the estimated data symbol according to the different respective values of DMAN for the different nodes.

Preferably the receiver is a digital mobile radio.

According to a third aspect of the present invention there is provided a digital communications system including an RBF network for the recovery of data symbols from received signals, characterised in that the RBF network is arranged to provide joint detection of SDMA (spatial division multiple access) datastreams.

The use of RBF for recovery of SDMA signals makes possible a significant increase in the effective capacity of the communications system. This aspect of the present invention may advantageously be used, for example, in the context of GSM cellular telephone systems, since this system meets the requirements on mobile frequency error control necessary to support joint demodulation.

Systems embodying the present invention will now be described in further detail, by way of example only, with reference to the accompanying drawings: in which Figure 1 is a block diagram of a digital communications system; Figures 2a and 2b are schematics of alternative RBF network topologies: Figure 3 is a diagram showing an RBF network for use in a system embodying the present invention; and Figure 4 is plot of BER simulation results for a two element diversity combiner.

A digital communications system includes a transmitter 1 comprising an encoder 2 and an RF TX stage 3. The data to be transmitted is input to the encoder which produces a symbol stream using, e.g., 4-state PSK (phase-shift keying), which is output via the TX stage onto the communications channel. At a receiver 7, the radio signal is detected by an RF front-end 4 and applied to an RBF network 5. As received, the signal will typically have been modified by the transmission characteristics of the communications channel and will include noise and possibly cross-talk from other signals on the channel. The RBF network 5 produces from this noisy signal estimates of the transmitted symbol value, and these symbol values are passed to the decoder 6 which recreates the data stream.

In the examples described below and in the appendix, the RBF 5 is implemented in dedicated CMOS VLSI logic while the associated control processes, and in particular the adaptive algorithm used to provide appropriate values for the network node centres, are implemented in software in an embedded DSP such as a Texas Instruments TMS series device.

The DSP may be a fixed or floating point device. These implementations are described by way of example only.

Alternative implementations may, for example, dispense with the DSP and implement the adaptive algorithm as part of the CMOS circuits used for the RBF.

The communications system, which is described in simplified form above, may in practice be, for example, a digital cellular telephone system such as GSM.

Figures 2a and 2b show two possible topologies for the RBF network, and other topologies are possible. The RBF network may be combined with other techniques such as decision feedback equalisation (DFE), linear transversal equalisation (LTE) and linear beamforming. The key advantage of the RBF approach over currently used methods is its improved bit error rate (BER) performance in channels which are limited by time dispersion (inter-symbol interference). This allows the bit rate to be increased (all other parameters being equal) in a given radio propagation environment.

An RBF diversity combiner has a superior BER performance to a linear combiner because it is able to use the multipath components of the received signal constructively. An equaliser or diversity combiner using an RBF network is also less susceptible to the phenomenon of 'irreducible bit error rate' which effects existing receivers in channels with high time dispersion. When system performance is limited by time dispersion effects the RBF approach offers performance improvements which may be expressed in the following three equivalent statements: 1. For a given channel and data rate the number of array elements required to meet a target BER can be reduced. For example, simulation work carried out at ERA has demonstrated that under certain severe channel conditions a two element RBF diversity combiner out-performs a four element linear combiner.

2. For a given channel, data rate and number of array elements the achievable BER is improved. The improvement over linear techniques increases as the time dispersion and SNR are increased.

3. For a given channel, number of array elements and target BER the data rate can be increased.

Channel Equalisation In principle the RBF technique can be applied to channel equalisation in any digital communications system.

However, the most promising application area is digital mobile radio. The technique is probably not suitable for systems such as telephone modem applications because these systems use multi-level modulation techniques (eg 64-QAM) which would make the number of nodes required in the RBF network prohibitive. Radio systems tend to use modulation schemes with fewer states (typically 4 states as in OQPSK,GMSK) which makes the size of the RBF network manageable.

Diversity Combining Diversity combining may be regarded as a kind of 'spatial equalisation' in which the inputs to the RBF network are the signals from several independent receivers.

Diversity combining is widely used to reduce the effects of Doppler fading and time dispersion in digital mobile radio systems.

Since in both Channel Equalisation and Diversity Combining the RBF network inputs are digitised baseband signals a well designed RBF integrated circuit could be configured for either application.

Spatial Division Multiple Access (SDMA) RBF networks are suitable for performing joint detection of more than one user signal simultaneously.

Joint detection enables more than one user signal to share a physical channel (ie carrier frequency and time slot).

This offers the possibility of increasing the capacity of a cellular system. Joint detection can be performed on either temporal or spatial samples. In the latter case it is commonly referred to as Spatial Division Multiple Access (SDMA).

A particular candidate for this approach is the GSM cellular telephone system, since this system meets the requirements on mobile frequency error control necessary to support joint demodulation.

Figure 3 shows one example of an RBF network used in implementing the present invention. This is a modifed form of the RBF network which significantly reduces the complexity of a hardware implementation with little loss in performance. This has been achieved by replacing the traditional distance calculation in the network with an L1 distance measure (the Manhattan distance).

In a standard RBF network each node computes a function of the following form.

(4.1) 2 2.

Here aN is a constant and D2 is the squared Euclidean distance between the observation vector and the centre vector of the node, defined as follows (where Ne is the dimensionality of the observation vector, xi is the ith entry in the observation vector and ci is the ith entry in the state centre vector).

(4.2) This computation requires 2Ne real additions and 2Ne multiply-accumulate operations for each node in the network. For a diversity combiner the number of nodes required to cover all states of the observation vector is given by Ns Na (4.3) where Na is the number of states in the modulation scheme being used and L is the delay spread of the channel normalised to the symbol period. If joint detection is being performed the required number of nodes is squared.

The Gaussian exponential function of equation 4.1 can be realised using a look-up table. However the work at ERA has established that a high dynamic range is required in the output of this look up table if the overall BER performance of the network is not to be compromised. For example, for an Ne=8 dimensional (complex) input space an equivalent wordlength of 37 bits is required to sustain a BER of 10-5. In general a floating point representation would be required, with a floating point adder to combine the node outputs.

Using the method of the present invention, the computational cost associated with the distance computation can be significantly reduced by abandoning the use of the Euclidean distance measure and adopting instead the Manhattan distance measure, as defined below.

(4.4) The advantage of the Manhattan distance approach is the fact that no multiply operations are required in the computation, reducing the computational cost to 2Ne additions per node. Since a multiplier is a much more complex circuit than an adder this significantly reduces the circuit complexity and avoids the doubling of dynamic range associated with the squaring operation.

The work at ERA has also established that the computation of the Gaussian exponential function can be removed with negligible loss in BER performance. The reduced network, referred to as the Manhattan Distance-only network, has the form of a 'winner take all' network in which the node with the centre closest (in the Manhattan sense) to the observation sample is used for the output.

This is illustrated in Figure 3. The Ne dimensional complex data is effectively treated as 2Ne dimensional real data in this scheme. This architecture allows physical nodes to be time multiplexed to compute the outputs of several RBF nodes. The required number of physical nodes is given by Nu=(fsym/fc1k) 2NeNs (4.5) where Ns is the number of nodes in the RBF network being implemented, fsw is the symbol rate (the rate at which the network output must be computed) and fclk is the maximum clock rate of the node unit. A flexible architecture would allow the number of nodes Ns to be traded against the dimensionality of the input vector Ne for a given clock and data rate.

For example, assuming fCtk=64 MHz, fS=l MBaud/s and No=8, a total of 64 physical nodes would be needed to implement a network of Ns=256 nodes. This number of nodes would be sufficient for handling (maximum) delay spreads of up to four symbol periods with a fourstate modulation scheme.

The BER performance of this network has been investigated and compared against a conventional RBF network by computer simulation. The results obtained from one particular simulation scenario are given in Figure 4.

This simulation used an Ne=2 element diversity combiner.

The BER estimates are averaged over an ensemble of 1000 random static channels. The channels are generated assuming an exponential delay profile with an rms delay spread (normalised to the symbol period) of 0.5. The channels to the two array elements are assumed to be uncorrelated. BPSK modulation with square root raised cosine filtering is used, and 10,000 bits are used to estimate the BER in each channel.

The points marked 'o' show the performance of a double precision floating point simulation of the conventional RBF network. The points marked 'x' show the performance of a fixed point Manhattan Distance-only implementation where the input samples are quantised to 8 bits. Also shown are the results obtained with the optimum linear (Wiener MMSE) diversity combiner.

Systems embodying the present invention use a modified form of the RBF function in which all multiply operations have been eliminated resulting in an architecture suitable for digital VLSI implementation. The computational load has been reduced to 2Ne addition operations per node where Ne is the dimensionality of the (complex) input space.

Simulation results show that the performance loss in terms of BER compared to the conventional RBF network is negligible. This reduction in cost and complexity opens up a number of new application areas for RBF processors, in particular that of digital mobile radio. In such systems the RBF processor could be employed for channel equalisation, diversity combining or to increase the capacity of the system by means of joint detection (SDMA).

The appendix below gives further implementation details and describes alternative embodiments.

APPENDIX 1. RBF NETWORK The architecture of an RBF network is shown below.

Radial Basis Function Network 112 Sean ed Data x1(k) C2 xjk) I I Decsion (t)imat Data X2(k) symbol xk) Ȧi;s v Observation Euclidean Gaussian Vector X(k) Distance Nodes Computation Figure 1.1 = Radial Basis Function Network The input layer consists of a number of (complex valued) signals which together make up the observation vector X(k) at time k X(k)=[x1(k) x2(k)...xN(k)]T The observation vector may consist of digital samples from a tapped delay line or samples from an array of diversity receivers. This observation vector is distributed in parallel to a number Ns of hidden layer nodes. Each node computes the Euclidean distance between the observation vector and a centre vector belonging to that node

where dlfk) is the distance computation output of the ith node and Ci is the (complex valued) centre vector of the ith node ci =kij C,2 CiN Each distance measure di(k) is applied to a processor with a Gaussian transfer function, which computes the quantities

where K and o are constant parameters of the network. The Ns quantities gi(k) are forwarded to the decision device. The decision device sums groups of the node outputs together to produce a set of N decision variables.

There will be one such decision variable for every possible data symbol produced by the source. For example, if BPSK modulation is used Na=2, whereas for GMSK modulation Na=4. The source symbol associated with the largest decision variable vm(k) is output as the estimated transmitted symbol a(k).

2. MANHATTAN DISTANCE ONLY RBF NETWORK In the Manhattan distance only RBF network the observation vector X(k), the centre vectors Ci and the number of nodes Ns are defined as for the standard RBF network. The Euclidean distance calculation performed by each node is replaced by the Manhattan distance calculation as defined below.

In the Manhattan distance only network there are no Gaussian elements and the Ns distance metrics are passed straight to the decision device. The decision device finds the node with the smallest distance metric and outputs the source symbol associated with that node as the estimated transmitted symbol a(k).

Manhattan Ci DistancbOnly h QBF Nwork x1(k) 02 Xk) ^- I stimated Data Device } symbolEstimated Data symbol xkk) sum boy Observation Manhattan Vector X(k) Distance Computation Figure 2.1 : Manhattan Distance only RBF Network 3. NETWORK ARCHITECTURES FOR EQUALISATION AND DIVERSITY COMBINING An RBF network can be used to perform channel equalisation as illustrated below. The input to the RBF network can be successive samples of the input signal (Transversal Equaliser) or successive samples of the output signal (Decision Feedback Equaliser) or a combination of the two.

MDRBF ~ a(k) a(k) Transversal Equaliser Transversal Equaliser with Decision Feedback Figure 3.1 = RBF Networks for Channel Equalisation RBF networks can also be applied to diversity combining. Some possible architectures are illustrated below. The inputs to the network can come from the array elements directly (Narrowband Diversity Combiner) or from tapped delay lines behind each array element (Wideband Diversity Combiner). Diversity combining can also be combined with Decision Feedback.

The RBF network can also be applied to the output of a linear combiner (conventional adaptive antenna). The linear combiner applies different weights to each array antenna signal before summing them. This produces a 1 dimensional (complex) signal which can be postprocessed by an RBF network. This architecture may be particularly important in digital radio systems subject to co-channel interference. The linear combiner is effective at suppressing co-channel interference signals, and can therefore improve the c/I of the received signal before it is applied to the RBF network for demodulation.

MD RBF I1 > w A a(k) MDRBFa(k) r1 r1 Narrowband Diversity Combiner Wideband Diversity Combiner MD MD RBF t (k) < A MDRBF r1 w2 w3 Narrowband Diversity Linear Diversity Combiner Combiner with with RBF Post Processor Decision Feedback Figure 3.2: RBF Networks for Diversity Combining 4. RBF COMPONENTS IN THE CONTEXT OF DIGITAL RADIO RECEIVERS The figure below shows how an RBF network could be employed as an equaliser in a digital radio transceiver such as a GSM handset. The signal from the receiving antenna is amplified with a Low Noise Amplifier before being applied to a quadrature downconverter. This downconverter extracts the I and Q components of the received signals and produces two baseband outputs. The baseband signals are digitised in separate A/D converters resulting in a quadrature baseband digital signal. Samples of this signal enter a tapped delay line. The sample rate at this point is typically equal to the bit rate although higher sampling rates could also be used. The samples from the tapped delay line are input to the RBF network which outputs an estimate of the transmitted data bit at time k. The resulting bit stream is forwarded to the speech CODEC.

Meprr'e t- I Trnng > D e JL 7; crtce ADC LNA L ;J1 a(k) A 11 , a(k) Frequency and Timing Error t; > Control Quadrature Downconversion Figure 4.1 : RBF Equaliser in the Context of a Digital Radio Receiver (eg GSM handset transceiver) An adaptive algorithm is required to control the RBF network by providing appropriate values for the network node centres Cj. The algorithm would typically be implemented in software to compute the centres from the observation data in parallel to the RBF network itself. The adaptive algorithm may be a channel estimation algorithm as currently employed in GSM or a clustering algorithm such as is commonly employed in controlling RBF networks. In either case such algorithms typically rely on a training sequence of known data symbols embedded in the received data, or by feeding back the decisions produced by the RBF network into the adaptive processor.

No restriction should be placed on the type of algorithm used to control the RBF network since this patent is concerned with the implementation of the RBF network itself, given a set of node centres. This operation may be considered separate from the means used in the computation of the node centres.

The next figure illustrates how an RBF network could be used in an SDMA array diversity receiver, such as could be applied to a GSM basestation. The signals from a number of array elements are amplified, quadrature downconverted and digitised. The digital signals are applied to an RBF network. It is assumed that the received signals consist of two user signals with additive noise. The RBF processor performs joint detection of the two user signals and outputs two bit streams. Once again an adaptive processor is required to compute the network centres.

Quadrature Downconversion 7 MD I /7 - MD RBF LNA e + . A a(k) user 1 I o CODEC > -f > - LIDC CODEC ~ ~ ~ < ' Us rtTtaning Algorfttwn Sequence - ~ L LJ User 2Training Sequence Figure 4.2 : RBF SDMA Diversity Combiner in the Context of a Digital Radio Receiver (eg GSM basestation transceiver)

Claims (7)

1. CLAIMS 1. A method of operating a digital communications system including applying a noisy signal to a radial basis function (RBF) network comprising a plurality of nodes corresponding to different symbol values and outputting from the network an estimated data symbol, characterised in that the nodes of the network compute a parameter DMAN, where

   and in that the network selects the value of the estimated data symbol according to the different respective values of for for the different nodes.
2. 2. A method according to claim 1, in which the input to the network comprises successive samples of a signal from a single receiver.
3. 3. A method according to claim 2, in which the input to the network comprises simultaneous samples from an array of receivers.
4. 4. A method according to any one of the preceding claims including joint detection of transmitted data symbols from more than one source.
5. 5. A receiver for a digital communications system including a radial basis function (RBF) network comprising a plurality of nodes corresponding to different symbol values and arranged to output from the network an estimated data symbol, characterised in that the nodes of the network are arranged to compute a parameter DMN, where

   and in that the network selects the value of the estimated data symbol according to the different respective values of for for the different nodes.
6. 6. A digital mobile radio including a receiver according to claim 5.
7. 7. A digital communications system including an RBF network for the recovery of data symbols from received signals, characterised in that the RBF network is arranged to provide joint detection of SDMA (spatial division multiple access) datastreams.