GB2351885A

GB2351885A - Neural network for real-time channel equalisation

Info

Publication number: GB2351885A
Application number: GB9914894A
Authority: GB
Inventors: Gordon Shiells; Hamish Grant; Tim Spracklen
Original assignee: Axeon Ltd
Current assignee: Johnson Matthey Battery Systems Engineering Ltd
Priority date: 1999-06-26
Filing date: 1999-06-26
Publication date: 2001-01-10
Also published as: GB9914894D0; AU5552100A; WO2001001344A3; WO2001001344A2

Abstract

A Neural Network capable of extracting digital data from a radio signal has a sufficiently high degree of parallelism to dynamically determine at least one suitable channel model presenting a selected propagation path of the radio signal. The degree of parallelism provided by the Neural Network is sufficient to obtain channel equalisation of the received signal by processing data inputted into the Neural Network as the data are received in real-time from the radio signal.

Description

2351885 1 2 3 4 Neural Network for performing real-time Channel

Equalisation.

6 7 This invention relates to an Artificial Neural Network 8 (ANN) implementation of Channel Equalisation, in 9 particular to a method of Channel Equalisation using a dedicated neural network chip. More particularly, it 11 relates to the use of an ANN to estimate a channel 12 model between a transmitter and a wireless receiver, 13 especially to a channel model between a base station 14 and a portable or mobile handset. It is particularly, but not exclusively, relevant for channel equalisation 16 for multi-media communication signals, e.g. voice, 17 data, and/or video signals. Even more particularly, 18 the invention relates to channel equalisation'in a 19 Universal Mobile Telephony System (UMTS) utilising an ANN included in a third generation telephone handset.

21 22 Modern mobile communication systems are becoming 23 ubiquitous. The penetration of second generation GSM 24 mobile telephones in the UK has reached over 25% and third generation systems are now being developed. The 26 European third generation systems - UMTS - provide high 27 speed data connectivity at 2 Mbits/s, to enable hand 28 held terminals and phones to support multimedia, e 29 mail, video conferencing as well as traditional voice and fax services. UMTS is based on the GSM protocol 31 and a migration path from GSM to UMTS exists. Further, 2 1 UMTS terminals are backwards compatible with GSM as 2 well as some proposed interim higher speed solutions 3 such as 'Edge' and the General Packet Radio Service 4 GPRS. 5 6 To enable mobile devices to operate at the speeds 7 demanded under UMTS, particularly in the indoor 8 environment, several technical problems must be 9 mitigated or obviated. One important technical problem 10 which must be dealt with relates to a mobile receiver 11 receiving at least one 'echo, of a radio signal. For 12 example, signals which propagate through an indoors 13 environment undergo reflections from any electrically 14 conducting objects they encounter (for example, any 15 metals). In any normal indoors environment therefore 16 several objects can exist which will act as 'mirrors' 17 for radio signals and a mobile receiver operating in 18 such an environment will therefore receive multiple 19 'echoes' of a radio signal. In addition, a mobile 20 telephone may have to operate in a noisy electrical 21 environment, for example with PCs, fluorescent lights, 22 electrical motors, televisions, and other mobile 23 telecommunication devices all capable of creating 24 interference. 25 26 Practical implementations to enable third generation 27 systems to handle the higher data rates envisioned 28 under UMTS currently fall short of their theoretical 29 ideals. Improving practical implementations towards 30 their optimum attainable state is likely to be a 31 crucial factor in determining the success of third 32 generation systems. In particular, the process of 33 demodulating and decoding the radio signals received 34 and effectively and efficiently extracting the digital 35 data contained therein affects the practicality and 36 cost effectiveness of these systems.

3 1 In order to effectively and efficiently convert the 2 received radio signals in a UMTS handset into a digital 3 bit stream, several operations need to be carried out. 4 5 Normal packet-based communication consists of data 6 being spread across several packets, which contain a 7 header signifying various items of information 8 including details of the packet type and routing 9 related information. In GSM, channel equalization is 10 carried out by examining a series of bits indicating 11 the GSM standard Imidamble' within the data section of 12 every packet of information sent from a base station to 13 a mobile handset. For each packet it receives, a 14 handset utilises a predetermined Imidamblel, which it 15 compares with the received signal. An appropriate 16 channel model is then determined by comparing the 17 received signal known midamble bits. 18 19 By assuming all the bits in the packet have been 20 affected in an identical manner the model of the 21 channel can be used to estimate the value of the data 22 bits (i.e. not the midamble bits) in the packet. 23 24 This process (called channel equalization') is carried 25 out for every packet received and imposes a significant 26 computational burden on the receiver. It should be 27 noted that this process makes no use of immediately 28 preceding estimates of the channel model; rather it 29 recomputes the channel model for every packet 30 individually. 31 32 Existing UMTS channel equalizers only make very 33 approximate estimates of the channel model, this is 34 mainly as a result of the additional imposed 35 computational burden and as a result of the difficulty 36 in handling the increased inter-symbol Interference 4 (ISI) which is described in more detail hereinbelow.

For example, given a 200-fold increase in bit rate, a 4 pro rata increase is generated in the computational load. Moreover, given such an increase in the bit 6 rate, any distortion inherent in the channel may be 7 exaggerated.

8 9 In communications terminology, the effect of errors can be measured by the Bit Error Rate (BER) which is a 11 measure of the number of bits in error in the decoded 12 data stream. A typical BER for a local area network 13 (LAN) such as Ethernet is 1 in 10'0. This means that 1 14 in every 10" bits will be in error. A typical satellite link will have a BER of 1 in 103 to 105. A 16 mobile telephone link will have a BER that can be as 17 poor as 1 in 101. In other words, the mobile link can 18 be a billion times worse than a normal computer network 19 in terms of the number of errors experienced. Clearly 20 sophisticated signal processing techniques are required 21 if such a poor quality link is to carry reliable data 22 for speech and computer applications. 23 24 Achieving channel equalization in an indoor environment 25 is particularly problematic at high bit rates. 26 Reflections generate multipath propagation of the 27 signal. A handset must be capable of selecting an 28 appropriate channel from several multiple copies of the 29 received signal. The received signal consists of 30 "echoes" of the actual signal dispersed in time due to 31 the variation in path-length. Each signal delayed by a 32 small additional time as a result of the slightly 33 longer distance travelled by these reflected signals 34 gives rise to what is known as Intersymbol Interference 35 (ISI) as Fig. 1 illustrates. 36 1 The received signal is thus a summation of many 2 independent transmitted,echo,, signals, each relatively 3 delayed in time and suffering differing attenuation. 4 At any given position or point, at a particular 5 frequency component of the modulated signal, the time6 delayed echo signals, will add vectorially. 7 Obviously, any movement of the receiver (or transmitter 8 or reflecting surfaces) will affect this vector 9 summation. 10 11 Fig 1 includes a diagrammatic sketch illustrating the 12 effect of time dispersion in the frequency domain. A 13 large delay spread in the received signals potential 14 gives rise to a higher BER than is the case when the is delay spread is small. If the delay spread becomes of 16 the order of the time per transmitted symbol, then the 17 BER will be dramatically affected. For example, if the 18 delay spread is of the order of say 30% of a 19 transmitted symbol period, ISI increases as the 20 received signal deteriorates. This is illustrated in 21 the lower sketch of Fig. 1. 22 23 The illustrated signal degradation shown in Fig. 1 24 ignores the effect of amplitude and phase variation. 25 In practice, the amplitude and phase of each of the 26 many delayed components of the signal will undergo 27 independent random variation resulting in an 28 irreducible mean bit-error ratio (BER). This BER 29 depends on the delay spread of the received components, 30 the period of the transmitted symbol and the sampling 31 time, and is roughly proportional to the square of the 32 normalised RMS delay spread (normalised to the symbol 33 period). The RMS delay spread experienced is thus 34 affected by the operating environment and, for example, 35 typically ranges at present from around 50 ns to 300 ns 36 in indoor environments.

6 1 UMTS will implement a much higher data rate than is 2 currently the case. This, together with a shorter 3 symbol period, gives rise to the potential for much 4 larger delay spreads to occur in a signal received in 5 an indoor environment than is presently encountered. 6 The RMS delay range can therefore be much larger than 7 the range given above for present data transmission 8 rates, and third-generation handsets will require 9 computational means capable of discriminating between 10 the channels received in order to reduce the delay 11 spread and ISI at the higher data rates envisioned 12 under UMTS. 13 14 To recapitulate, the multi-path delay depends on the 15 signal propagation environment. However, the increased 16 data transfer rates under UMTS could result, for 17 example, in bit rates being faster by factors of 200 18 than the current GSM bit rates. The multi-path delay, 19 when measured as a proportion of the bit time, will 20 thus increase proportionately; i.e., under UMTS, this 21 could similarly increase by a factor of 200. This will 22 greatly exaggerate the effects of ISI as such levels of 23 delay spread can produce bits which may not merely 24 overlap but which are shifted entirely out of their 25 assigned time slot as is illustrated in Fig. 1. 26 27 Most data communication either presently in GSM or 28 under UMTS will involve data transmission through what 29 is usually a relatively noisy channel. Upon reception, 30 the data are decoded and processed by the receiving 31 handset. 32 33 The receiver must determine a suitable channel model 34 representing the propagation path of the transmitted 35 signal. Selecting the best channel model should enable 36 the receiver to eliminate multi-path "echo" effects.

7 1 Channel Equalisation is a method of constructively 2 recombining the time-spread signals. Conventionally, 3 channel equalisation is carried out using digital 4 signal processing and errors made in estimating the 5 channel model by the handset will reduce its ability to 6 correctly decode the received data. That is, the BER 7 observed will be higher than would be the case if a 8 better channel model were to be used. Currently, the 9 channel models used in GSM and UMTS are relatively 10 poor, limiting their effectiveness in calculating the 11 degradation of the BER that is experienced solely as a 12 result of the channel model used. Consequently, even a 13 UMTS handset with a perfect channel model could have a 14 non-zero BER as a result of channel noise. is 16 The invention seeks to obviate or mitigate the 17 aforementioned problems in the prior art by providing 18 an Artificial Neural Network (ANN) to implement Channel 19 Equalisation. 20 21 ANNs are primarily sophisticated pattern classifiers 22 capable of extracting information from noisy or 23 uncertain environments. An ANN is a computational 24 paradigm loosely based on the biological neural model. 25 The majority (though not all) research into ANNs has 26 been undertaken using computer simulations based on PC 27 or Unix platforms. ANNs are effectively pattern 28 classifiers that have the capability to operate in 29 noisy or fuzzy, environments. Traditional statistical 30 pattern classifiers are closely related to ANNs but 31 ANNs have the ability to dynamically adapt to different 32 input data sets and do not require programming in the 33 conventional sense. 34 35 In order to respond appropriately to any given input, 36 conventional ANNs require prior training on example

8 1 data. Such training enables the ANN to be able to 2 adapt or converge on statistically significant patterns 3 in the input data set. Being probabilistic devices 4 each neuron in an ANN is initialized with a series of 5 random 'weights, and, as a result, the convergence on 6 the patterns in the input data is purely one of 7 statistical chance. For example, with a particular 8 distribution of initial weights, on one occasion the 9 ANN may converge on a particular pattern in the input 10 data set, on another occasion - with a different weight 11 distribution - this pattern may be missed'. 12 13 The patterns used to train the ANN should be typical, 14 of the patterns that will be observed in the real data 15 used during the execution' phase of operation. The 16 quality, (relevance) of the training data will have a 17 major impact on the capability of the ANN during 18 execution. If the ANN has not been trained on data 19 containing examples of patterns that are of interest, 20 then it will be unable to identify such patterns in the 21 'execution' data. Additionally, like biological neural 22 systems, the ANN must be shown many examples of the 23 training data - training runs containing hundreds of 24 thousands of cycles are not at all uncommon. 25 26 There are typically two different methods used for 27 training - supervised and unsupervised. In supervised 28 training, the input data are fed to the ANN while the 29 'known, answer is provided to the network. The network 30 then attempts to adjust its internal weights so as to 31 best represent the input data. 32 33 ANNs that implement a Self-Organising Map (SOM) can 34 utilise the structure of data received to dynamically 35 "train" upon as the data are received i.e. a SOM is 3G capable of unsupervised training under real-time 9 1 conditions. The SOM ANN attempts to form internal 2 classifications of significant clusters of data 3 'observed, in data they are input. 4 5 The invention seeks to use an ANN to extract digital 6 data from a communications signal in a noisy 7 environment. The communications signal is received 8 wirelessly. 9 10 In particular, the invention relates to an ANN 11 procedure for channel equalization', which computes a 12 model of the radio channel between a base station and a 13 handset and subsequently uses this model to estimate 14 the corruption experienced by the radio signal. 15 16 The present invention therefore seeks to provide a 17 process by which the estimation of the channel model is 18 undertaken in an entirely different manner than 19 heretofore known in the art in which the channel 20 equalization previously performed by a conventional 21 signal processor is now performed by an Artificial 22 Neural Network (ANN). 23 24 According to a first aspect of the invention there is 25 provided a Neural Network which is capable of 26 extracting digital data from a wireless signal input to 27 the Neural Network. The Neural Network includes input 28 means to receive said signal and has a Self-Organising 29 Map having a degree of parallelism which enables the 30 Neural Network to be capable of determining at least 31 one suitable channel model for the input signal. The 32 degree of parallelism is sufficiently high to perform 33 channel equalisation in real-time by processing the 34 digital data as the signal is input into the Neural 35 Network. 36 1 Preferably, the Self-Organising Map is a Modular Map. 2 The degree of parallelism is dependent on the bit rate 3 of the digital data of the signal. 4 5 Preferably, the Neural Network is implemented on a 6 dedicated chip. Preferably, the dedicated chip has the 7 aspects and embodiments which are described in the 8 attached Appendix which includes the text of GB 9 9902115.6. 10 11 Preferably, the Neural Network includes a neuron 12 comprising an arithmetic logic unit; a shifter 13 mechanism; a set of registers; an input port; an output 14 port; and control logic. 15 16 Preferably, the Neural Network includes a module 17 controller for controlling the operation of said at 18 least one neuron, the controller comprising an input 19 port; an output port; a programmable read-only memory; 20 an address map; an input buffer; and at least one 21 handshake mechanism. 22 23 Preferably, the Neural Network includes a neuron 24 module, the module comprising at least one neuron; and 25 at least one module controller. The at least one neuron 26 and the at least one module controller may be 27 implemented on one device. The at least one neuron and 28 the at least one module controller may be implemented 29 on at least one device taken from the group consisting 30 of: a full-custom very large scale integration (VLSI) 31 device, a semi-custom VLSI, and an application specific 32 integrated circuit (ASIC). 33 34 Preferably, at least two neuron modules are coupled 35 together. The neuron modules may be coupled in a 36 lateral expansion mode, alternatively, the neuron 11 1 modules may be coupled in a hierarchical mode. The 2 number of neurons in a module may be a power of two, 3 preferably 256. 4 5 The said at least one handshake mechanism may include a 6 synchronisation handshake mechanism for synchronising 7 data transfer between a sender and a receiver module. 8 9 Preferably, the synchronisation handshake mechanism 10 comprises a three-line mechanism which has three states 11 comprising a wait state, a no device state and a data 12 ready state and wherein the three-line mechanism 13 comprises two outputs from the receiver and one output 14 from the sender so that point-to-point data 15 transmission can occur without the need for additional 16 devices between the sender and receiver or vice versa. 17 18 Preferably, the Neural Network is incorporated into a 19 signal receiver. More preferably, the receiver is 20 incorporated into a handset. 21 22 Preferably, the wireless signal input occupies at least 23 part of the radio portion of the electromagnetic 24 spectrum. The radio signal may include at least one 25 from the group consisting of voice, data and video type 26 signals. 27 28 Preferably, the bit rate of the digital data of the 29 input signal is at least equal to or greater than 1 30 Mbits/s. More preferably, the bit rate of the digital 31 data of the input signal may be substantially in the 32 range 1 Mbit/s to 10 Mbits/s. 33 34 According to a second aspect of the invention, there is 35 provided a Neural Network channel equalizer which 36 includes a Neural Network according to the first aspect 12 1 of the invention.

2 3 According to a third aspect of the invention, there is 4 provided a handset for use in a wireless environment, the handset being capable of receiving the wireless 6 signal and including a Neural Network channel equalizer 7 according to the second aspect of the invention.

8 9 According to a fourth aspect of the invention, there is provided a method of training a neural network to 11 extract digital data from a wireless signal input to 12 the Neural Network, the method comprising the steps of 13 providing a network of neurons; reading an input vector 14 applied to the input of the neural network; calculating is the distance between the input vector and a reference 16 vector for all neurons in the network; finding the 17 active neuron; outputting the location of the active 18 neuron; and updating the reference vectors for all 19 neurons in a neighbourhood around the active neuron.

21 Preferably, the method of training a Neural Network 22 trains the Neural Network to function as a channel 23 equalizer according to the second aspect of the 24 invention.

26 According to a fifth aspect of the invention, there is 27 provided a method of channel equalization, in a 28 wireless signal environment, comprising the steps of:

29 using the Neural Network according to the first aspect of the invention to compute at least one channel model 31 representing a signal path between a base station and a 32 handset; using the computed channel model to determine 33 a level of distortion in the signal received by said 34 handset; and reconstructing the signal received by the handset to reduce the level of distortion in the 36 reconstructed signal.

13 1 Preferably, the reduction of the level of distortion in 2 the reconstructed signal reduces the Inter-Symbol 3 Interference of the received signal.

4 Embodiments of the present invention will now be 6 described, by way of example only, with reference to 7 the accompanying drawings, in which:

8 9 Fig. 1 is a sketch illustrating Inter-Symbol Interference between two spread out components of a 11 transmitted signal and the resultant degradation in the 12 received signal; and 13 14 Fig. 2 is a sketch which illustrates diagrammatically is an example of two different propagation paths in an 16 indoor environment a signal has taken before being 17 received by a handset (RX).

18 19 In a conventional digital handset, channel equalisation is performed by constructively recombining received 21 time-spread signals using digital signal processing 22 (DSP). The DSP operations carried out are based upon 23 certain assumptions about the channel model which are 24 derived from theoretical considerations relating to radio wave propagation in various environments.

26 27 For example, radio waves may mirror off both known and 28 unknown structural objects in a building 1, such as 29 Fig. 2 attempts to exemplify. The resulting difference in propagation times from the transmitter to 31 the receiver 2 (RX) give rise to a time spread in the 32 signal a receiver 2 receives, such as Fig. 1 33 illustrates.

34 Several known channel models have been formulated 36 according to current radio propagation theory. These 14 1 channel models' attempt to represent a communication 2 channel as a series of mathematical equations. These 3 equations are used to provide the necessary 4 equalization for the mobile handset by solving them for the known mid-amble bits. The channel models so 6 determined therefore represent a compromise between 7 computational efficiency and physical accuracy.

8 9 Radio transmission in an indoors environment requires a high degree of computational efficiency when performing 11 channel equalisation if an acceptable channel model is 12 to be obtained. Three-dimensional multi-path 13 propagation affords a relatively high number of degrees 14 of freedom that enable efficient channel selection to is be achieved using suitable optimisation processes 16 providing appropriate computational effort can be 17 expended.

18 19 Self-organised anarchic allocation processes have been proposed in the past for Dynamic Channel Allocation but 21 the practical implementation of such processes is 22 relatively difficult. Such processes rely on a handset 23 being capable of receiving all channels and selecting 24 an appropriate channel on the basis of carrier-to interference ratios. A handset must be capable of 26 operating in a co-channel interference limited 27 condition and needs to incorporate means for 28 appropriately performing channel equalisation at the 29 data rates received. DCA on the basis of the least, or the most acceptable level of interference needs to be 31 performed continuously and is highly intensive 32 computationally.

33 34 obviously, interference levels may change, and a channel initially selected may subsequently 36 deteriorate, for example, as a person using the handset moves around. To compensate for this, known channel 2 equalisation algorithms make no prior assumptions 3 regarding the optimum channel selected and do not 4 incorporate a "history" of the previously selected channel model.

6 7 Conventional hardware ANN implementations are mainly 8 either based on conventional DSP hardware (i.e. a 9 conventional DSP chip programmed to behave like an ANN) or contain a very limited number of neurons. As a 11 result, these hardware ANN implementations are not 12 suitable for DCA applications in a mobile telephony 13 environment.

14 one embodiment of the invention provides an ANN that 16 implements a Modular Map as a basic building block in 17 its SOM. The Modular Map design is such that many 18 modules can be connected together to create a wide 19 variety of configurations and network sizes, which enables a scalable system to be constructed. Such 21 systems can thus be adapted to cope with increased 22 parallelism and so mitigates the requirement for an 23 extensive increase in training time, such as is 24 associated, for example, with unitary implementations.

26 One embodiment of the invention incorporates a 27 dedicated NN chip architecture into a handset or 28 terminal capable of receiving signals in a radio 29 telephony environment. The NN chip includes a Modular Map (MM) design in its SOM, such as the related UK 31 Patent Application GB 9902115.6 describes, which 32 provides the high degree of parallelism required for an 33 ANN to constructively recombine time-spread signals in 34 high bit rate environments. The NN chip thus provides a NN channel equalizer which is capable of operating in 36 high bit rate environments such as, for example, the 16 1 UMTS environment. 2 3 The text of GB 9902115.6 filed herewith as the Appendix 4 is hereby deemed to be incorporated into the body of 5 the description. 6 7 The degree of parallelism the MM ANN chip provides is 8 capable of computing channel models with a level of 9 complexity heretofore unattainable. The MM ANN chip 10 provides a sufficiently adaptive parallelism to 11 efficiently construct suitable channel models under the 12 conditions imposed by UMTS. Thus even in conditions 13 where the data rate is high, for example, 2Mbits/sec, 14 superior channel models than those known at present can 15 be constructed. Such superior channel models can 16 provide for "fade-out" of signals in indoors 17 environments rather than "cut-out" as the reconstructed 18 signal BER gradually increases rather than suddenly 19 jumps. 20 21 The ANN can be initially trained on a wide variety of 22 radio propagation data - taken from actual measurements 23 made using real mobile handsets. For a digital 24 handset, the indoor environment - with its multiple 25 reflecting surfaces - presents the most difficult 26 problems such as, for example, Fig. 2 illustrates. 27 However, by incorporating the MM ANN chip into a 28 handset receiver, the reception qualities of a digital 29 handset are greatly improved compared with those 30 envisioned by present technology under high bit rate 31 conditions. 32 33 In one embodiment of the invention, the MM ANN was 34 trained using a wide variety of practical measurements 35 representing the received and non-equalized signal and 36 the known correct data. The training process is

17 1 essentially that followed conventionally when training 2 a Neural Network. The ANN is able to classify the 3 relationship between the received and known correct 4 data, and adjusts its internal weights in the 5 conventional manner so that it effectively learns, by 6 experience. One particular advantage of training the 7 ANN to act as a channel equaliser in a high bit rate 8 environment compared to a conventional channel 9 equaliser is the ability of the ANN able to take 10 advantage implicitly that the channel model does not 11 change rapidly with respect to time in terms of the bit 12 period, even in an indoor environment. The fact that a 13 channel model will not change too rapidly compared with 14 the received bit-rate in an indoor environment is well 15 known and can be observed in practical measurements. 16 Conventional current channel equalizers nonetheless 17 recompute the channel model completely for each data 18 packet, even when receiving data at high bit rates in 19 an indoor environment. 20 21 In effect, the ANN can be said to start with the 22 knowledge of the existing channel model and merely 23 computes the difference between the previous and 24 present channel modes. 25 26 A receiver according to one aspect of the invention 27 including an NN chip according to the appendix as a 28 channel equaliser effectively exploits this ability of 29 the ANN to incorporate its knowledge of an existing 30 channel model to compute the difference between a 31 previous and a present channel model. The degree of 32 parallelism of the SOM of the ANN enables the ANN to 33 respond to slight changes in the propagation 34 environment of the received signal even at high bit 35 rates and so optimise the channel model it computes. 36 18 1 Using conventional indoor propagation data available, 2 the ANN is trained over a wide variety ofconditions. 3 The initial training of an ANN is laboratory based. 4 The ANN is exposed to a variety of channel models and 5 propagation environments in order to predetermine the 6 relevant weights which are then fixed. These weights 7 are then incorporated into the architecture of a SOM of 8 an NN chip according to an embodiment of the attached 9 appendix. 10 11 The NN chip is then manufactured with these weights 12 predetermined so that the chip is capable of 13 functioning as a channel equalizer according to a 14 second aspect of the invention. The predetermined is weights incorporated on the NN chip thus represent the 16 experience gained by the laboratory-based parent ANN in 17 the laboratory training process. Such predetermined 18 weights are only representative of the variety of 19 conditions under which the laboratory ANN model was 20 trained. 21 22 It is well known that existing channel models are 23 deficient and are merely approximations to the observed 24 channel response. In one embodiment of the invention, 25 the ANN is trained on real data which improves the 26 channel model obtained. By training the ANN on real 27 data, the ANN thus provides a better estimate of the 28 actual channel behaviour than those obtained from 29 present theoretical and idealized mathematical channel 30 models. 31 32 Conventional in parallel GSM and UMTS channel 33 equalization algorithms can be used in order to provide 34 a comparison with the ANN-based solution. Under multi35 path propagation conditions and/or at high bit rates, 36 the ANN-based channel equaliser consistently out- 19 1 performs conventional DSP channel equalisers which 2 incorporate conventional channel models. 3 4 In another embodiment of the invention, the ANN is 5 allowed to continue training into the execution phase 6 and, as a result, the ANN is able to track variations 7 in the propagation environment that may not have been 8 experienced during the dedicated training (and, of 9 course, which may not be present in a mathematical 10 channel model). Laboratory training can, at best, be 11 representative of only a subset of the total 12 environmental conditions that a receiver will 13 experience in use. By allowing the NN to adapt the 14 weights predetermined under laboratory conditions to is incorporate local, experience gained in actual use of 16 the receiver incorporating the NN in the field, channel 17 equalisation can be improved. Incorporating such 18 'local' experience to optimise the channel model 19 parameters has not heretofore been possible using 20 conventional DSP channel equalizers. The NN, being 21 adaptive, will quickly learn' about new propagation 22 conditions, even those not represented in conventional 23 channel models. 24 25 This embodiment of the ANN MM provides an average 26 processing gain of 3dB. Its performance is 27 particularly effective during poor conditions when the 28 conventional equalizer was unable to provide a 29 competent estimate of the channel model. 30 31 A conventional channel equalizer suffers from a 32 catastrophic or non-graceful degradation under 33 worsening signal conditions. That is, as the 34 propagation conditions worsen, the BER will likewise 35 worsen but, at some critical point, the BER will 36 collapse. This is frequently observed on conventional

1 GSM handsets in the sense that the handset is said to 2 "work or notwork" - i.e. it "cuts-out" rather than 3 "fades-out" with the move between the two conditions 4 being highly sensitive to worsening propagation 5 conditions. 6 7 In one embodiment of the invention, the ANN solution 8 was found not only to have an average 3dB processing 9 gain, but also to exhibit a gradual degradation of 10 reception quality under worsening signal propagation 11 conditions - i.e the signal "fades-out". Furthermore 12 it provides a better performance under poor conditions 13 when a conventional DSP approach is unable to operate. 14 15 16 This is achieved using a simulation of the current 17 (prototype) version of the Modular Map chip at a clock 18 rate of 50MHz. Commercial versions are expected to 19 operate at 200MHz and an improvement in performance is 20 expected. 21 22 An ANN channel equaliser therefore provides several 23 advantages over conventional DSP channel equalisers. 24 By incorporating a dedicated NN chip according to the 25 appendix into a receiver of a mobile telecommunications 26 device, such as for example, a mobile phone or suitable 27 computer device, superior signal reception can be 28 obtained compared to that provided by conventional DSP 29 means at bit rates. 30 31 Moreover, the NN chip provides further advantages, for 32 example, low power consumption. 33 34 Furthermore, the degree of parallelism of the NN chip 35 can be adapted according to the anticipated bit rate. 36 In particular, the NN's ability to provide superior 21 1 channel models and to reduce ISI in a received signal 2 enables more reliable communication to be effected. 3 The reliability of data transfers is improved over a 4 wider range of propagation conditions than convention 5 DSP means are able to cope with: any deterioration with 6 time of the channel models determined by the NN is 7 gradual rather than stepped. This provides for gradual 8 variations in the BER of data transfers rather than 9 sudden rises. 10 utilising the best channel model is particularly 12 important in multi-media type communications which are 13 less error tolerant. For example, video-type data has 14 less redundancy than voice-type data and thus has a 15 lower error tolerance. The NN chip channel equaliser 16 thus provides a means to effect more reliable multi 17 media type data communication in a high bit rate 18 environment and under a range of propagation 19 environments, particularly but not exclusively, in an 20 indoor environment. 21 22 While several embodiments of the present invention have 23 been described and illustrated, it will be apparent to 24 those skilled in the art once given this disclosure 25 that various modifications, changes, improvements and 26 variations may be made without departing from the 27 spirit or scope of this invention. 28 29 other variations on the invention include, for example, 30 varying the initial ANN training to predetermine 31 weights for the subsequent NN chip manufacture. For 32 example, the training may be carried out using a 33 conventional ANN or by using a prototype NN chip. 34 35 The NN channel equalizer may be directly included in a 36 receiver component or provided as a separate component,

22 1 for example in a handset or portable terminal. 2 References to a 'handset, are considered to include 3 references to any device capable of receiving a 4 wireless communication signal. For example, reference 5 to a handset is deemed to include a reference to a 6 mobile telecommunications device which is held in the 7 hand of a user, e.g. a mobile telephone or terminal. 8 Furthermore, a reference to a handset is considered to 9 also include a reference to any another mobile or 10 wireless device such as, for example, a portable data 11 communicator, a portable modem or modem-type card 12 whether capable of being connected to a computer-type 13 device or actually included in a computer-type device, 14 especially, for example, a PCMCIA card. The size of is the communications device is not deemed to be limiting 16 providing the device utilises wireless communications. 17 Reference to a wireless communications device is 18 considered to include referring to remote controls for 19 appliances and apparatus, and wireless headphones etc. 20 21 The NN chip is especially suited to channel 22 equalisation of delay spread signals in high bit-rate 23 environments; i.e. bit rates in excess of 2 Mbit/s, for 24 example, 8 Mbit/s. However, channel equalisation can 25 be similarly provided at lower bit rates and the above 26 figures are not intended to provide strict limits. The 27 NN chip can be adapted to suit the transmission 28 conditions to consistently eliminate ISI in digital 29 environments. 30 31 Under UMTS the spectrum range is in the radio part of 32 the electromagnetic spectrum. For other applications 33 the spectrum range of the transmission media could 34 differ; for example the spectrum range of the signals 35 could include infra-red, optical, Ultra-violet, 36 millimetric radio, microwave radio, and/or VHF/UHF 23 1 millimetric radio, microwave radio, and/or VHF/UHF 2 radio signals. Further more, the spectrum range can 3 include the audio and near-audio range, for example 4 ultrasonic frequencies, especially 1 to 15 MHz, and voice frequencies, especially 300 to 3400 MHz.

24

Claims

1 Claims:

2 3 1. A Neural Network which is capable of extracting 4 digital data from a wireless signal input"to the Neural Network, the Neural Network including input means to 6 receive said signal; the Neural Network having:

7 a Self-Organising Map having a degree of 8 parallelism which enables the Neural Network to be 9 capable of determining at least one suitable channel model for the input signal, wherein the degree of 11 parallelism is sufficiently high to perform channel 12 equalisation in real-time by processing the digital 13 data as the signal is input into the Neural Network.

14

2. A Neural Network as claimed in claim 1, wherein 16 the Self-Organising Map is a Modular Map.

17 18

3. A Neural Network as claimed in either Claim 1 or 19 Claim 2, wherein the degree of parallelism is dependent on the bit rate of the digital data of the signal.

21 22

4. A Neural Network as claimed in any preceding 23 Claim, wherein the Neural Network is implemented on a 24 dedicated chip.

26

5. A Neural Network as claimed in any one of Claims 1 27 to 4, wherein the Neural Network includes a neuron 28 comprising:

29 an arithmetic logic unit; a shifter mechanism; 31 a set of registers; 32 an input port; 33 an output port; and 34 control logic.

36

6. A Neural Network as claimed in any one of Claims 1 1 to 5, wherein the Neural Network includes a module 2 controller for controlling the operation of said at 3 least one neuron, the controller comprising 4 an input port; an output port; 6 a programmable read-only memory; 7 an address map; 8 an input buffer; and 9 at least one handshake mechanism.

11

7. A Neural Network as claimed in any one of Claims 5 12 or 6, wherein the Neural Network includes a neuron 13 module, the module comprising 14 at least one neuron; and at least one module controller.

16 17.

8. A Neural Network as claimed in any one of Claims 5 18 to 7, wherein the at least one neuron and the at least 19 one module controller are implemented on one device.

21

9. A Neural Network as claimed in any one of Claims 5 22 to 8, wherein the at least one neuron and the at least 23 one module controller are implemented at least one 24 device taken from the group consisting of: a full custom very large scale integration (VLSI) device, a 26 semi-custom VLSI, and an application specific 27 integrated circuit (ASIC).

28 29

10. A Neural Network as claimed in any one of Claims 5 to 9 comprising 31 at least two neuron modules coupled together.

32 33

11. A Neural Network as claimed in claim 10, wherein 34 the neuron modules are coupled in a lateral expansion mode.

36 26 1

12. A Neural Network as claimed in claim 10 or 11, 2 wherein the neuron modules are coupled in a 3 hierarchical mode. 4 5

13. A Neural Network as claimed in any one of Claims 5 6 to 12, wherein the number of neurons in a module is a 7 power of two. 8 9

14. A Neural Network as claimed in Claim 13, wherein 10 the number of neurons in a module is 256. 11 12

15. A Neural Network as claimed in any one of Claims 5 13 to 14, wherein said at least one handshake mechanism 14 includes a synchronisation handshake mechanism for 15 synchronising data transfer between a sender and a 16 receiver module. 17 18

16. A Neural Network as claimed in Claim 15, wherein 19 the synchronisation handshake mechanism comprises a 20 three- line mechanism which has three states comprising 21 a wait state, a no device state and a data ready state 22 and wherein the three-line mechanism comprises two 23 outputs from the receiver and one output from the 24 sender so that point-to-point data transmission can 25 occur without the need for additional devices between 26 the sender and receiver or vice versa. 27 28

17. A Neural Network as claimed in any preceding 29 claim, wherein the Neural Network is incorporated into 30 a receiver. 31 32

18. A Neural Network as claimed in Claim 17, wherein 33 the receiver is incorporated into a wireless 34 communications device. 35 36

19.A Neural Network as claimed in any preceding 27 1 Claim, wherein the wireless signal input occupies at 2 least part of the radio portion of the electromagnetic 3 spectrum.

4

20. A Neural Network as claimed in Claim 19, wherein 6 the radio signal includes at least one from the group 7 consisting of voice, data and video type signals.

8 9

21. A Neural Network as claimed in Claim 19, wherein the bit rate of the digital data of the input signal is 11 substantially in the range 1 Mbit/s to 10 Mbits/s.

12 13 21. A Neural Network channel equalizer which includes 14 a Neural Network as claimed in any one preceding claim.

is 16

22. A handset for use in a wireless environment, the 17 handset being capable of receiving the wireless signal 18 and including Neural Network channel equalizer as 19 claimed in claim 21.

21

23. A method of training a Neural Network as claimed 22 in any one of Claims 1 to 20 to extract digital data 23 from a wireless signal input to the Neural Network, the 24 method comprising the steps of providing a network of neurons; 26 reading an input vector applied to the input of 27 the neural network; 28 calculating the distance between the input vector 29 and a reference vector for all neurons in the network; finding the active neuron; 31 outputting the location of the active neuron; and 32 updating the reference vectors for all neurons in 33 a neighbourhood around the active neuron.

34 36

24.A method of training a Neural Network as claimed 28 1 in Claim 23, wherein the method trains the Neural 2 Network to function as a channel equalizer as claimed 3 in Claim 21.

4

25. A method of channel equalization in a wireless 6 signal environment, comprising the steps of:

7 using the Neural Network as claimed in any one of 8 claims 1 to 21 to compute at least one channel model 9 representing a signal path between a base station and a handset.

11 using the computed channel model to determine a 12 level of distortion in the signal received by said 13 handset; and 14 reconstructing the signal received by the handset to reduce the level of distortion in the reconstructed 16 signal.

17 18

26. A method of channel equalisation as claimed in 19 claim 25, wherein the reduction of the level of distortion in the reconstructed signal reduces the 21 Inter-Symbol Interference of the received signal.

22 23

27. A Neural Network for extracting information from a 24 radio signal substantially as described herein and with reference to the accompanying drawings.

26 27

28. A Neural Network channel equalizer for extracting 28 information from a radio signal substantially as 29 described herein and with reference to the accompanying drawings.

31 32

29. A method of channel equalisation using a 33 Neural Network substantially as described herein and 34 with reference to the accompanying drawings.

36